m3 engine parser (reasoning-only) stage

Signed-off-by: Ting Sun <suntcrick@gmail.com>

.buildkite/test-amd.yaml

@@ -647,7 +647,7 @@ steps:
   - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
 
 - label: e2e Core (1 GPU) # TBD
-  timeout_in_minutes: 180
+  timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
   agent_pool: mi250_1
   optional: true
@@ -2075,19 +2075,6 @@ steps:
   - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
   - pytest -v -s v1/spec_decode/test_acceptance_length.py -m slow_test
 
-- label: e2e Core (1 GPU) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/v1/
-  - tests/v1/e2e/
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
-
 - label: e2e Scheduling (1 GPU) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]

.buildkite/test_areas/engine.yaml

@@ -74,6 +74,16 @@ steps:
     - tests/v1/e2e/general/
   commands:
     - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
+  mirror:
+    amd:
+      device: mi250_1
+      timeout_in_minutes: 35
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/
+      - tests/v1/e2e/general/
+      - vllm/platforms/rocm.py
 
 - label: V1 e2e (2 GPUs)
   key: v1-e2e-2-gpus

.buildkite/test_areas/lm_eval.yaml

@@ -109,6 +109,7 @@ steps:
       - image-build-amd
       commands:
       - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - export PYTORCH_ROCM_ARCH=gfx942 # Limit Quark compilation to save time
       - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx.txt
 
 - label: MoE Refactor Integration Test (H100 - TEMPORARY)

csrc/libtorch_stable/attention/mla/sm100_cutlass_mla_kernel.cu

@@ -268,9 +268,14 @@ int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_ba
   using TileShapeD = typename MlaSm100Type::TileShapeD;
   arguments.problem_shape =
       cute::make_tuple(TileShapeH{}, static_cast<int>(max_seq_len), TileShapeD{}, static_cast<int>(num_batches));
-  // Assumes device 0 when getting sm_count.
-  arguments.hw_info.sm_count =
-      sm_count <= 0 ? cutlass::KernelHardwareInfo::query_device_multiprocessor_count(/*device_id=*/0) : sm_count;
+  if (sm_count <= 0) {
+    int current_device = 0;
+    cudaGetDevice(&current_device);
+    arguments.hw_info.sm_count =
+        cutlass::KernelHardwareInfo::query_device_multiprocessor_count(current_device);
+  } else {
+    arguments.hw_info.sm_count = sm_count;
+  }
   arguments.split_kv = static_cast<int>(num_kv_splits);
   MlaSm100Type::Fmha::set_split_kv(arguments);
 

csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu

@@ -301,8 +301,9 @@ __global__ void per_token_group_quant_8bit_packed_register_kernel(
 
   const int sf_k_local = local_group_id % kGroupsPerBlockX;
   const int row_local = local_group_id / kGroupsPerBlockX;
-  const int sf_k_idx = blockIdx.x * kGroupsPerBlockX + sf_k_local;
-  const int mn_idx = blockIdx.y * kRowsPerBlock + row_local;
+  // Rows on grid.x: mn scales with tokens and can exceed the 65535 grid.y cap.
+  const int sf_k_idx = blockIdx.y * kGroupsPerBlockX + sf_k_local;
+  const int mn_idx = blockIdx.x * kRowsPerBlock + row_local;
 
 #if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
   asm volatile("griddepcontrol.wait;");
@@ -496,14 +497,15 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
                   " is not a multiple of 4.");
   const int kx = GetGroupsPerBlockX(padded_groups_per_row);
   const int ry = 16 / kx;
-  const int64_t blocks_x = padded_groups_per_row / kx;
-  const int64_t blocks_y = (tma_aligned_mn + ry - 1) / ry;
+  const int64_t row_blocks = (tma_aligned_mn + ry - 1) / ry;
+  const int64_t sf_k_blocks = padded_groups_per_row / kx;
   const int num_threads = (kx * ry) * THREADS_PER_GROUP;
-  // CUDA caps grid.x and grid.y at 2^31 - 1; guard against pathological inputs.
-  STD_TORCH_CHECK(blocks_x <= static_cast<int64_t>(INT32_MAX) &&
-                      blocks_y <= static_cast<int64_t>(INT32_MAX),
+  // CUDA caps grid.x at 2^31 - 1 and grid.y at 2^16 - 1 (65535).
+  constexpr int64_t kMaxGridDimYZ = 65535;
+  STD_TORCH_CHECK(row_blocks <= static_cast<int64_t>(INT32_MAX) &&
+                      sf_k_blocks <= kMaxGridDimYZ,
                   "per_token_group_quant_8bit_packed grid too large: (",
-                  blocks_x, ", ", blocks_y, ").");
+                  row_blocks, ", ", sf_k_blocks, ").");
 
   auto dst_type = output_q.scalar_type();
 
@@ -513,8 +515,8 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
   #define LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, KX, RY)                         \
     do {                                                                       \
       cudaLaunchConfig_t config = {};                                          \
-      config.gridDim = dim3(static_cast<unsigned int>(blocks_x),               \
-                            static_cast<unsigned int>(blocks_y));              \
+      config.gridDim = dim3(static_cast<unsigned int>(row_blocks),             \
+                            static_cast<unsigned int>(sf_k_blocks));           \
       config.blockDim = dim3(num_threads);                                     \
       config.dynamicSmemBytes = 0;                                             \
       config.stream = stream;                                                  \
@@ -539,8 +541,8 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
 #else
   #define LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, KX, RY)                         \
     do {                                                                       \
-      dim3 grid(static_cast<unsigned int>(blocks_x),                           \
-                static_cast<unsigned int>(blocks_y));                          \
+      dim3 grid(static_cast<unsigned int>(row_blocks),                         \
+                static_cast<unsigned int>(sf_k_blocks));                       \
       dim3 block(num_threads);                                                 \
       per_token_group_quant_8bit_packed_register_kernel<T, DST_DTYPE, 128, KX, \
                                                         RY>                    \

tests/engine/test_arg_utils.py

@@ -649,3 +649,196 @@ def test_cloud_storage_tokenizer_skips_get_model_path(monkeypatch):
     args = EngineArgs(model="s3://bucket/model", tokenizer="s3://bucket/tokenizer")
     assert args.model == "s3://bucket/model"
     assert args.tokenizer == "s3://bucket/tokenizer"
+
+
+class TestDeviceIds:
+    def test_device_ids_with_cvd_out_of_range(self, monkeypatch):
+        """--device-ids index beyond the CVD set raises ValueError."""
+        from vllm.platforms import current_platform
+
+        key = current_platform.device_control_env_var
+        monkeypatch.setenv(key, "4,5")
+        args = EngineArgs(model="m", device_ids=[0, 2])
+        with pytest.raises(ValueError, match="out of range"):
+            args._resolve_device_ids()
+
+    def test_device_ids_with_cvd_resolve_to_physical_ids(self, monkeypatch):
+        """--device-ids are CVD-local indices resolved to physical ids."""
+        from vllm.platforms import current_platform
+
+        key = current_platform.device_control_env_var
+        monkeypatch.setenv(key, "4,5")
+        args = EngineArgs(model="m", device_ids=[0, 1])
+        assert args._resolve_device_ids() == [4, 5]
+
+    def test_device_ids_with_uuid_cvd_resolve_to_physical_ids(self, monkeypatch):
+        """--device-ids support UUID CVD values resolved by the platform."""
+        from vllm.platforms import current_platform
+
+        key = current_platform.device_control_env_var
+        monkeypatch.setenv(key, "GPU-abcd1234,GPU-ef567890")
+        monkeypatch.setattr(
+            type(current_platform),
+            "device_control_id_to_physical_device_id",
+            classmethod(
+                lambda cls, device_id: {"GPU-abcd1234": 4, "GPU-ef567890": 5}[device_id]
+            ),
+        )
+
+        args = EngineArgs(model="m", device_ids=[0, 1])
+        assert args._resolve_device_ids() == [4, 5]
+
+    def test_device_ids_with_uuid_args_resolve_to_physical_ids(self, monkeypatch):
+        """UUID --device-ids are resolved to physical IDs immediately."""
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(
+            type(current_platform),
+            "device_control_id_to_physical_device_id",
+            classmethod(lambda cls, device_id: {"GPU-abcd1234": 4}[device_id]),
+        )
+
+        args = EngineArgs(model="m", device_ids=["GPU-abcd1234"])
+        assert args._resolve_device_ids() == [4]
+
+    def test_device_ids_reject_mixed_integer_and_uuid_args(self):
+        """--device-ids must not mix CVD indices and UUIDs."""
+        args = EngineArgs(model="m", device_ids=[0, "GPU-abcd1234"])
+        with pytest.raises(ValueError, match="must not mix"):
+            args._resolve_device_ids()
+
+    def test_no_device_ids(self):
+        """No --device-ids returns None."""
+        args = EngineArgs(model="m")
+        assert args._resolve_device_ids() is None
+
+    def test_cli_parsing(self):
+        """--device-ids parses comma-separated string from CLI."""
+        parser = FlexibleArgumentParser()
+        EngineArgs.add_cli_args(parser)
+        parsed = parser.parse_args(["--model", "m", "--device-ids", "0,2,4"])
+        assert parsed.device_ids == [0, 2, 4]
+
+    def test_cli_parsing_uuid(self):
+        """--device-ids parses comma-separated UUID strings from CLI."""
+        parser = FlexibleArgumentParser()
+        EngineArgs.add_cli_args(parser)
+        parsed = parser.parse_args(
+            ["--model", "m", "--device-ids", "GPU-abcd1234,GPU-ef567890"]
+        )
+        assert parsed.device_ids == ["GPU-abcd1234", "GPU-ef567890"]
+
+    def test_assigned_physical_gpu_ids_are_physical_with_cvd(self, monkeypatch):
+        """assigned_physical_gpu_ids are already physical and not composed with CVD."""
+        import vllm.platforms.interface as platform_interface
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(platform_interface, "_assigned_physical_gpu_ids", [4, 5])
+        monkeypatch.setenv(current_platform.device_control_env_var, "4,5")
+
+        assert current_platform.device_id_to_physical_device_id(0) == 4
+        assert current_platform.device_id_to_physical_device_id(1) == 5
+        assert current_platform.logical_device_id_to_visible_device_id(0) == 0
+        assert current_platform.logical_device_id_to_visible_device_id(1) == 1
+
+    def test_assigned_physical_gpu_ids_map_to_visible_uuid_cvd(self, monkeypatch):
+        """Physical IDs map back to visible ordinals when CVD uses UUIDs."""
+        import vllm.platforms.interface as platform_interface
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(platform_interface, "_assigned_physical_gpu_ids", [5])
+        monkeypatch.setenv(
+            current_platform.device_control_env_var,
+            "GPU-abcd1234,GPU-ef567890",
+        )
+        monkeypatch.setattr(
+            type(current_platform),
+            "device_control_id_to_physical_device_id",
+            classmethod(
+                lambda cls, device_id: {"GPU-abcd1234": 4, "GPU-ef567890": 5}[device_id]
+            ),
+        )
+
+        assert current_platform.logical_device_id_to_visible_device_id(0) == 1
+
+    def test_device_ids_reject_duplicates(self):
+        """--device-ids must not contain duplicate entries."""
+        args = EngineArgs(model="m", device_ids=[2, 2])
+        with pytest.raises(ValueError, match="duplicates"):
+            args._resolve_device_ids()
+
+    def test_cli_parsing_strips_whitespace(self):
+        """--device-ids tolerates whitespace around commas."""
+        parser = FlexibleArgumentParser()
+        EngineArgs.add_cli_args(parser)
+        parsed = parser.parse_args(["--model", "m", "--device-ids", "0, 2, 4"])
+        assert parsed.device_ids == [0, 2, 4]
+
+    def test_visible_ordinal_to_physical_ignores_assigned_ids(self, monkeypatch):
+        """visible_device_id_to_physical_device_id maps torch device ordinals,
+        independent of the logical-to-physical mapping.
+
+        Regression test: CustomAllreduce passes device.index (a visible
+        ordinal) and must not index into assigned_physical_gpu_ids, which
+        raised IndexError for non-identity --device-ids like [2, 3].
+        """
+        import vllm.platforms.interface as platform_interface
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(platform_interface, "_assigned_physical_gpu_ids", [2, 3])
+        monkeypatch.delenv(current_platform.device_control_env_var, raising=False)
+
+        # CVD unset: visible ordinal == physical ID, even beyond the
+        # assigned list's length.
+        assert current_platform.visible_device_id_to_physical_device_id(2) == 2
+        assert current_platform.visible_device_id_to_physical_device_id(3) == 3
+
+        monkeypatch.setenv(current_platform.device_control_env_var, "4,5")
+        assert current_platform.visible_device_id_to_physical_device_id(1) == 5
+        with pytest.raises(IndexError, match="out of range"):
+            current_platform.visible_device_id_to_physical_device_id(2)
+
+
+class TestDpDeviceIdSharding:
+    def test_dp_supervisor_device_ids_stay_env_relative(self):
+        """Regression test: the DP supervisor must pass env-relative indices,
+        not physical IDs, because each child re-resolves --device-ids
+        against its inherited device-control env var."""
+        import argparse
+
+        from vllm.entrypoints.openai.dp_supervisor import _build_device_ids
+
+        args = argparse.Namespace(
+            tensor_parallel_size=2, pipeline_parallel_size=1, device_ids=None
+        )
+        assert _build_device_ids(args, local_rank=0) == [0, 1]
+        assert _build_device_ids(args, local_rank=1) == [2, 3]
+
+    def test_dp_supervisor_shards_user_device_ids(self):
+        """User-provided --device-ids are sharded across DP children."""
+        import argparse
+
+        from vllm.entrypoints.openai.dp_supervisor import _build_device_ids
+
+        args = argparse.Namespace(
+            tensor_parallel_size=2, pipeline_parallel_size=1, device_ids=[4, 5, 6, 7]
+        )
+        assert _build_device_ids(args, local_rank=0) == [4, 5]
+        assert _build_device_ids(args, local_rank=1) == [6, 7]
+        with pytest.raises(ValueError, match="needs devices"):
+            _build_device_ids(args, local_rank=2)
+
+    def test_dp_rank_shards_user_assigned_gpu_ids(self):
+        """get_physical_gpu_ids_for_local_dp_rank slices the user-provided
+        --device-ids list instead of recomputing from the env var."""
+        from vllm.platforms import current_platform
+        from vllm.v1.engine.utils import get_physical_gpu_ids_for_local_dp_rank
+
+        evar = current_platform.device_control_env_var
+        assert get_physical_gpu_ids_for_local_dp_rank(
+            evar, local_dp_rank=1, world_size=2, user_assigned_gpu_ids=[4, 5, 6, 7]
+        ) == [6, 7]
+        with pytest.raises(ValueError, match="needs devices"):
+            get_physical_gpu_ids_for_local_dp_rank(
+                evar, local_dp_rank=2, world_size=2, user_assigned_gpu_ids=[4, 5, 6, 7]
+            )

tests/entrypoints/anthropic/test_anthropic_messages_conversion.py

@@ -8,6 +8,8 @@ AnthropicServingMessages._convert_anthropic_to_openai_request().
 Also covers extended-thinking edge cases such as ``redacted_thinking``
 blocks echoed back by Anthropic clients, and streaming conversion in
 ``message_stream_converter``.
+
+Also covers cache usage computation in ``_build_anthropic_usage``.
 """
 
 import json
@@ -18,7 +20,11 @@ import pytest
 from vllm.entrypoints.anthropic.protocol import (
     AnthropicMessagesRequest,
 )
-from vllm.entrypoints.anthropic.serving import AnthropicServingMessages
+from vllm.entrypoints.anthropic.serving import (
+    AnthropicServingMessages,
+    _build_anthropic_usage,
+    _get_cached_tokens,
+)
 from vllm.entrypoints.openai.chat_completion.protocol import (
     ChatCompletionResponseStreamChoice,
     ChatCompletionStreamResponse,
@@ -27,6 +33,7 @@ from vllm.entrypoints.openai.engine.protocol import (
     DeltaFunctionCall,
     DeltaMessage,
     DeltaToolCall,
+    PromptTokenUsageInfo,
     UsageInfo,
 )
 
@@ -653,6 +660,108 @@ class TestThinkingBlockConversion:
         assert asst.get("content") == "Hi!"
 
 
+# ======================================================================
+# Cache usage computation
+# ======================================================================
+
+
+class TestGetCachedTokens:
+    """Tests for _get_cached_tokens helper."""
+
+    def test_none_usage(self):
+        assert _get_cached_tokens(None) is None
+
+    def test_no_prompt_tokens_details(self):
+        usage = UsageInfo(prompt_tokens=100, completion_tokens=10)
+        assert _get_cached_tokens(usage) is None
+
+    def test_cached_tokens_present(self):
+        usage = UsageInfo(
+            prompt_tokens=100,
+            completion_tokens=10,
+            prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=80),
+        )
+        assert _get_cached_tokens(usage) == 80
+
+    def test_cached_tokens_zero(self):
+        """Zero cached tokens should return 0, not None."""
+        usage = UsageInfo(
+            prompt_tokens=100,
+            completion_tokens=10,
+            prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=0),
+        )
+        assert _get_cached_tokens(usage) == 0
+
+    def test_cached_tokens_none_in_details(self):
+        usage = UsageInfo(
+            prompt_tokens=100,
+            completion_tokens=10,
+            prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=None),
+        )
+        assert _get_cached_tokens(usage) is None
+
+
+class TestBuildAnthropicUsage:
+    """Tests for _build_anthropic_usage helper.
+
+    Anthropic defines: total_input = input_tokens + cache_read + cache_creation
+    vLLM's prompt_tokens is the total.
+    """
+
+    def test_no_cache_info(self):
+        """When cache info is unavailable, return raw prompt_tokens."""
+        result = _build_anthropic_usage(100, 10, None)
+        assert result.input_tokens == 100
+        assert result.output_tokens == 10
+        assert result.cache_read_input_tokens is None
+        assert result.cache_creation_input_tokens is None
+
+    def test_cache_hit(self):
+        """When cache is hit, input_tokens excludes cached tokens."""
+        usage = UsageInfo(
+            prompt_tokens=100,
+            completion_tokens=10,
+            prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=80),
+        )
+        result = _build_anthropic_usage(100, 10, usage)
+        assert result.input_tokens == 20  # 100 - 80
+        assert result.output_tokens == 10
+        assert result.cache_read_input_tokens == 80
+        assert result.cache_creation_input_tokens == 0
+
+    def test_zero_cached_tokens(self):
+        """Zero cached tokens should still set cache_creation to 0."""
+        usage = UsageInfo(
+            prompt_tokens=100,
+            completion_tokens=10,
+            prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=0),
+        )
+        result = _build_anthropic_usage(100, 10, usage)
+        assert result.input_tokens == 100  # 100 - 0
+        assert result.cache_read_input_tokens == 0
+        assert result.cache_creation_input_tokens == 0
+
+    def test_all_tokens_cached(self):
+        """When all tokens are cached, input_tokens should be 0."""
+        usage = UsageInfo(
+            prompt_tokens=100,
+            completion_tokens=10,
+            prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=100),
+        )
+        result = _build_anthropic_usage(100, 10, usage)
+        assert result.input_tokens == 0
+        assert result.cache_read_input_tokens == 100
+        assert result.cache_creation_input_tokens == 0
+
+    def test_no_prompt_tokens_details(self):
+        """UsageInfo without prompt_tokens_details returns no cache info."""
+        usage = UsageInfo(prompt_tokens=100, completion_tokens=10)
+        result = _build_anthropic_usage(100, 10, usage)
+        assert result.input_tokens == 100
+        assert result.cache_read_input_tokens is None
+        assert result.cache_creation_input_tokens is None
+
+
 class TestInlineSystemMessageInMessagesArray:
     """Verify that ``role: system`` messages embedded inside the ``messages``
     array are preserved in their original position.
@@ -1098,6 +1207,135 @@ class TestMessageStartIncludesTypeAndRole:
         assert message["role"] == "assistant"
 
 
+class TestStreamingCacheUsageSemantics:
+    """Locks in the documented streaming behavior of cache usage fields.
+
+    vLLM's OpenAI chat completion streaming only attaches
+    ``prompt_tokens_details`` to the terminal usage chunk. The Anthropic layer
+    mirrors that contract: cache fields are omitted on ``message_start`` (key
+    absence signals "unknown") and populated on ``message_delta`` (the final
+    cumulative count). This is intentionally consistent with vLLM's OpenAI
+    behavior, even though Anthropic's upstream API populates cache fields on
+    ``message_start``; closing that gap requires plumbing cache info into the
+    first chunk at the OpenAI layer, which is out of scope here.
+    """
+
+    @pytest.mark.asyncio
+    async def test_streaming_cache_fields_absent_then_populated(self):
+        """First chunk lacks prompt_tokens_details (vLLM contract);
+        message_start omits cache fields. The final chunk carries
+        prompt_tokens_details, so message_delta carries resolved values."""
+
+        async def sse_input():
+            yield _make_stream_chunk(
+                delta=DeltaMessage(role="assistant", content="hi"),
+                usage=UsageInfo(prompt_tokens=100, total_tokens=100),
+            )
+            yield _make_stream_chunk(finish_reason="stop")
+            yield _make_stream_chunk(
+                choices=[],
+                usage=UsageInfo(
+                    prompt_tokens=100,
+                    completion_tokens=5,
+                    total_tokens=105,
+                    prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=80),
+                ),
+            )
+            yield "data: [DONE]"
+
+        converter = _make_stream_converter()
+        output = []
+        async for event in converter.message_stream_converter(sse_input()):
+            output.append(event)
+        events = _parse_sse_events(output)
+
+        # message_start: cache fields unknown → omitted from JSON entirely.
+        start_usage = events[0][1]["message"]["usage"]
+        assert events[0][0] == "message_start"
+        assert start_usage["input_tokens"] == 100
+        assert "cache_read_input_tokens" not in start_usage
+        assert "cache_creation_input_tokens" not in start_usage
+
+        # message_delta: authoritative usage with cache fields populated.
+        delta_usage = next(
+            data["usage"] for ev, data in events if ev == "message_delta"
+        )
+        assert delta_usage["input_tokens"] == 20  # 100 - 80
+        assert delta_usage["cache_read_input_tokens"] == 80
+        assert delta_usage["cache_creation_input_tokens"] == 0
+
+    @pytest.mark.asyncio
+    async def test_streaming_no_cache_hit(self):
+        """When the final chunk reports cached_tokens=0, message_delta carries
+        cache fields = 0 (cache miss); message_start still omits them."""
+
+        async def sse_input():
+            yield _make_stream_chunk(
+                delta=DeltaMessage(role="assistant"),
+                usage=UsageInfo(prompt_tokens=50, total_tokens=50),
+            )
+            yield _make_stream_chunk(finish_reason="stop")
+            yield _make_stream_chunk(
+                choices=[],
+                usage=UsageInfo(
+                    prompt_tokens=50,
+                    completion_tokens=5,
+                    total_tokens=55,
+                    prompt_tokens_details=PromptTokenUsageInfo(cached_tokens=0),
+                ),
+            )
+            yield "data: [DONE]"
+
+        converter = _make_stream_converter()
+        output = []
+        async for event in converter.message_stream_converter(sse_input()):
+            output.append(event)
+        events = _parse_sse_events(output)
+
+        start_usage = events[0][1]["message"]["usage"]
+        delta_usage = next(
+            data["usage"] for ev, data in events if ev == "message_delta"
+        )
+        assert start_usage["input_tokens"] == 50
+        assert "cache_read_input_tokens" not in start_usage
+        assert "cache_creation_input_tokens" not in start_usage
+        assert delta_usage["input_tokens"] == 50  # 50 - 0
+        assert delta_usage["cache_read_input_tokens"] == 0
+        assert delta_usage["cache_creation_input_tokens"] == 0
+
+    @pytest.mark.asyncio
+    async def test_streaming_no_prompt_tokens_details_at_all(self):
+        """If --enable-prompt-tokens-details is off, no chunk carries cache
+        info; both message_start and message_delta omit cache fields."""
+
+        async def sse_input():
+            yield _make_stream_chunk(
+                delta=DeltaMessage(role="assistant"),
+                usage=UsageInfo(prompt_tokens=30, total_tokens=30),
+            )
+            yield _make_stream_chunk(finish_reason="stop")
+            yield _make_stream_chunk(
+                choices=[],
+                usage=UsageInfo(prompt_tokens=30, completion_tokens=2, total_tokens=32),
+            )
+            yield "data: [DONE]"
+
+        converter = _make_stream_converter()
+        output = []
+        async for event in converter.message_stream_converter(sse_input()):
+            output.append(event)
+        events = _parse_sse_events(output)
+
+        start_usage = events[0][1]["message"]["usage"]
+        delta_usage = next(
+            data["usage"] for ev, data in events if ev == "message_delta"
+        )
+        assert "cache_read_input_tokens" not in start_usage
+        assert "cache_creation_input_tokens" not in start_usage
+        assert "cache_read_input_tokens" not in delta_usage
+        assert "cache_creation_input_tokens" not in delta_usage
+
+
 # ======================================================================
 # Auto-detection of system-first template requirement
 # ======================================================================

tests/entrypoints/openai/test_dp_supervisor.py

@@ -364,7 +364,7 @@ class MockVLLMServer:
         await self._serve_task
 
 
-def launch_mock_vllm(child_args: argparse.Namespace, env_updates: dict[str, str]):
+def launch_mock_vllm(child_args: argparse.Namespace):
     logger.info("Launching mock vLLM on port %s", child_args.port)
     mock_vllm = MockVLLMServer(
         port=child_args.port,
@@ -375,7 +375,7 @@ def launch_mock_vllm(child_args: argparse.Namespace, env_updates: dict[str, str]
 
 
 def launch_mock_vllm_with_drain(
-    child_args: argparse.Namespace, env_updates: dict[str, str]
+    child_args: argparse.Namespace,
 ):
     logger.info("Launching mock vLLM with 15s drain on port %s", child_args.port)
     mock_vllm = MockVLLMServer(

tests/entrypoints/serve/render/test_derender.py

@@ -8,6 +8,7 @@ import pytest
 import pytest_asyncio
 
 from tests.utils import RemoteLaunchRenderServer
+from vllm.tokenizers import get_tokenizer
 
 MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM"
 
@@ -486,3 +487,438 @@ async def test_derender_completion_kv_transfer_params_passthrough(client):
     )
     assert response.status_code == 200
     assert response.json()["kv_transfer_params"] == kv
+
+
+# ---------------------------------------------------------------------------
+# E2E: render -> derender roundtrip with parser (reasoning + tool calls)
+# ---------------------------------------------------------------------------
+
+PARSER_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
+
+_E2E_TOOLS = [
+    {
+        "type": "function",
+        "function": {
+            "name": "get_weather",
+            "description": "Get weather for a city",
+            "parameters": {
+                "type": "object",
+                "properties": {"city": {"type": "string"}},
+            },
+        },
+    }
+]
+
+
+@pytest.fixture(scope="module")
+def parser_server():
+    args = [
+        "--enable-auto-tool-choice",
+        "--tool-call-parser",
+        "hermes",
+        "--reasoning-parser",
+        "deepseek_r1",
+    ]
+    with RemoteLaunchRenderServer(PARSER_MODEL, args) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def parser_client(parser_server):
+    async with httpx.AsyncClient(
+        base_url=parser_server.url_for(""), timeout=60.0
+    ) as http_client:
+        yield http_client
+
+
+@pytest.fixture(scope="module")
+def parser_tokenizer():
+    return get_tokenizer(PARSER_MODEL)
+
+
+def _encode(tokenizer, text: str) -> list[int]:
+    return tokenizer.encode(text, add_special_tokens=False)
+
+
+def _decoded(tokenizer, token_ids: list[int]) -> str:
+    return tokenizer.decode(token_ids, skip_special_tokens=True)
+
+
+def _require_markers_survive(tokenizer, text: str, *markers: str) -> list[int]:
+    """Encode text and skip the test if any marker is lost in roundtrip."""
+    ids = _encode(tokenizer, text)
+    decoded = tokenizer.decode(ids, skip_special_tokens=False)
+    for m in markers:
+        if m not in decoded:
+            pytest.skip(f"Marker {m!r} lost in encode->decode roundtrip")
+    return ids
+
+
+async def _e2e_render_chat(
+    client: httpx.AsyncClient,
+    model: str,
+    messages: list[dict],
+) -> dict:
+    resp = await client.post(
+        "/v1/chat/completions/render",
+        json={"model": model, "messages": messages},
+    )
+    assert resp.status_code == 200, resp.text
+    return resp.json()
+
+
+def _e2e_generate_response(
+    token_ids: list[int],
+    request_id: str = "chatcmpl-e2e-test",
+) -> dict:
+    return {
+        "request_id": request_id,
+        "choices": [
+            {
+                "index": 0,
+                "token_ids": token_ids,
+                "finish_reason": "stop",
+            }
+        ],
+    }
+
+
+@pytest.mark.asyncio
+async def test_e2e_plain_roundtrip(parser_client, parser_tokenizer):
+    """Plain text without reasoning markers roundtrips correctly."""
+    messages = [{"role": "user", "content": "What is 2+2?"}]
+    gen_req = await _e2e_render_chat(parser_client, PARSER_MODEL, messages)
+
+    answer = "The answer is four."
+    output_ids = _encode(parser_tokenizer, answer)
+    expected = _decoded(parser_tokenizer, output_ids)
+
+    resp = await parser_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": PARSER_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    content = resp.json()["choices"][0]["message"]["content"]
+    assert content == expected
+
+
+@pytest.mark.asyncio
+async def test_e2e_token_identity(parser_client, parser_tokenizer):
+    """encode(derender(token_ids)) == token_ids (RL invariant)."""
+    messages = [{"role": "user", "content": "Hi"}]
+    gen_req = await _e2e_render_chat(parser_client, PARSER_MODEL, messages)
+
+    answer = "Hello! How can I help?"
+    output_ids = _encode(parser_tokenizer, answer)
+
+    resp = await parser_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": PARSER_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+        },
+    )
+    assert resp.status_code == 200
+    content = resp.json()["choices"][0]["message"]["content"]
+    re_encoded = _encode(parser_tokenizer, content)
+    assert output_ids == re_encoded
+
+
+@pytest.mark.asyncio
+async def test_e2e_non_ascii_roundtrip(parser_client, parser_tokenizer):
+    """CJK + emoji roundtrip without U+FFFD."""
+    messages = [{"role": "user", "content": "Reply in Chinese"}]
+    gen_req = await _e2e_render_chat(parser_client, PARSER_MODEL, messages)
+
+    answer = "你好世界 😀"
+    output_ids = _encode(parser_tokenizer, answer)
+
+    resp = await parser_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": PARSER_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+        },
+    )
+    assert resp.status_code == 200
+    content = resp.json()["choices"][0]["message"]["content"]
+    assert "�" not in content
+
+
+@pytest.mark.asyncio
+async def test_e2e_parsed_reasoning(parser_client, parser_tokenizer):
+    """<think>...</think> splits into reasoning + content."""
+    messages = [{"role": "user", "content": "What is 2+3?"}]
+    gen_req = await _e2e_render_chat(parser_client, PARSER_MODEL, messages)
+
+    reasoning_text = "The user wants 2 plus 3. That is 5."
+    answer_text = "The answer is 5."
+    output_text = f"<think>{reasoning_text}</think>{answer_text}"
+    output_ids = _require_markers_survive(parser_tokenizer, output_text, "</think>")
+
+    resp = await parser_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": PARSER_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+            "chat_request": {
+                "model": PARSER_MODEL,
+                "messages": messages,
+                "include_reasoning": True,
+            },
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    msg = resp.json()["choices"][0]["message"]
+    assert msg["reasoning"] is not None
+    assert reasoning_text in msg["reasoning"]
+    assert answer_text in msg["content"]
+    assert "<think>" not in msg["content"]
+
+
+@pytest.mark.asyncio
+async def test_e2e_parsed_tool_call(parser_client, parser_tokenizer):
+    """<tool_call> extracted into tool_calls field."""
+    messages = [{"role": "user", "content": "Weather in Paris?"}]
+    gen_req = await _e2e_render_chat(parser_client, PARSER_MODEL, messages)
+
+    output_text = (
+        "<think>Let me check the weather.</think>"
+        '<tool_call>\n{"name": "get_weather", '
+        '"arguments": {"city": "Paris"}}\n</tool_call>'
+    )
+    output_ids = _require_markers_survive(
+        parser_tokenizer,
+        output_text,
+        "</think>",
+        "<tool_call>",
+        "</tool_call>",
+    )
+
+    resp = await parser_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": PARSER_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+            "chat_request": {
+                "model": PARSER_MODEL,
+                "messages": messages,
+                "tools": _E2E_TOOLS,
+                "tool_choice": "auto",
+            },
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    choice = resp.json()["choices"][0]
+    assert choice["message"]["tool_calls"]
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "get_weather"
+
+
+@pytest.mark.asyncio
+async def test_e2e_parsed_reasoning_and_tool_call(parser_client, parser_tokenizer):
+    """Reasoning + tool call in the same output."""
+    messages = [{"role": "user", "content": "Weather in Paris?"}]
+    gen_req = await _e2e_render_chat(parser_client, PARSER_MODEL, messages)
+
+    reasoning_text = "I should look up the weather."
+    tool_text = (
+        '<tool_call>\n{"name": "get_weather", '
+        '"arguments": {"city": "Paris"}}\n</tool_call>'
+    )
+    output_text = f"<think>{reasoning_text}</think>{tool_text}"
+    output_ids = _require_markers_survive(
+        parser_tokenizer, output_text, "</think>", "<tool_call>"
+    )
+
+    resp = await parser_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": PARSER_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+            "chat_request": {
+                "model": PARSER_MODEL,
+                "messages": messages,
+                "tools": _E2E_TOOLS,
+                "tool_choice": "auto",
+                "include_reasoning": True,
+            },
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    choice = resp.json()["choices"][0]
+    assert choice["message"]["reasoning"] is not None
+    assert reasoning_text in choice["message"]["reasoning"]
+    assert choice["message"]["tool_calls"]
+
+
+@pytest.mark.asyncio
+async def test_e2e_no_chat_request_fallback(parser_client, parser_tokenizer):
+    """Without chat_request, derender falls back to plain detokenization."""
+    messages = [{"role": "user", "content": "Hello"}]
+    gen_req = await _e2e_render_chat(parser_client, PARSER_MODEL, messages)
+
+    answer = "Hi there!"
+    output_ids = _encode(parser_tokenizer, answer)
+
+    resp = await parser_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": PARSER_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+        },
+    )
+    assert resp.status_code == 200
+    content = resp.json()["choices"][0]["message"]["content"]
+    assert "Hi" in content
+
+
+# ---------------------------------------------------------------------------
+# E2E: HarmonyParser + GPT-OSS
+# ---------------------------------------------------------------------------
+
+HARMONY_MODEL = "openai/gpt-oss-20b"
+
+
+def _ensure_harmony_vocab():
+    """Pre-cache the o200k_base BPE file needed by openai-harmony.
+
+    The Rust tiktoken-rs backend downloads from Azure Blob Storage, which
+    may be unreachable in some environments.  When the cache is cold we
+    fetch the file ourselves and place it in ``/tmp/tiktoken-rs-cache/``
+    using the SHA-1(URL) filename that tiktoken-rs expects.
+    """
+    import hashlib
+    import urllib.request
+    from pathlib import Path
+
+    url = "https://openaipublic.blob.core.windows.net/encodings/o200k_base.tiktoken"
+    cache_dir = Path("/tmp/tiktoken-rs-cache")
+    cache_key = hashlib.sha1(url.encode()).hexdigest()
+    cache_file = cache_dir / cache_key
+    if not cache_file.exists():
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        urllib.request.urlretrieve(url, cache_file)
+
+
+@pytest.fixture(scope="module")
+def harmony_server():
+    _ensure_harmony_vocab()
+    args = [
+        "--trust-remote-code",
+        "--enable-auto-tool-choice",
+        "--tool-call-parser",
+        "openai",
+        "--reasoning-parser",
+        "openai_gptoss",
+    ]
+    with RemoteLaunchRenderServer(HARMONY_MODEL, args) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def harmony_client(harmony_server):
+    async with httpx.AsyncClient(
+        base_url=harmony_server.url_for(""), timeout=60.0
+    ) as http_client:
+        yield http_client
+
+
+@pytest.fixture(scope="module")
+def harmony_tokenizer():
+    return get_tokenizer(HARMONY_MODEL, trust_remote_code=True)
+
+
+def _harmony_extract_assistant_ids(
+    tokenizer, assistant_msg: dict, user_content: str = "test"
+) -> list[int]:
+    """Extract assistant token IDs via apply_chat_template diff."""
+    prompt = [{"role": "user", "content": user_content}]
+    full = prompt + [assistant_msg]
+    text_prompt = tokenizer.apply_chat_template(
+        prompt, add_generation_prompt=True, tokenize=False
+    )
+    text_full = tokenizer.apply_chat_template(
+        full, add_generation_prompt=False, tokenize=False
+    )
+    prompt_ids = tokenizer.encode(text_prompt)
+    full_ids = tokenizer.encode(text_full)
+    assistant_ids = list(full_ids[len(prompt_ids) :])
+    if not assistant_ids:
+        pytest.skip("Could not extract assistant tokens for Harmony")
+    return assistant_ids
+
+
+@pytest.mark.asyncio
+async def test_e2e_harmony_plain_roundtrip(harmony_client, harmony_tokenizer):
+    """GPT-OSS content-only roundtrip."""
+    messages = [{"role": "user", "content": "What is 2+2?"}]
+    gen_req = await _e2e_render_chat(harmony_client, HARMONY_MODEL, messages)
+
+    assistant_msg = {"role": "assistant", "content": "Four."}
+    output_ids = _harmony_extract_assistant_ids(harmony_tokenizer, assistant_msg)
+
+    resp = await harmony_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": HARMONY_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+            "chat_request": {
+                "model": HARMONY_MODEL,
+                "messages": messages,
+            },
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    content = resp.json()["choices"][0]["message"]["content"]
+    assert content is not None and len(content) > 0
+    assert "Four" in content
+
+
+@pytest.mark.asyncio
+async def test_e2e_harmony_reasoning(harmony_client, harmony_tokenizer):
+    """GPT-OSS reasoning: analysis channel extracted."""
+    messages = [{"role": "user", "content": "Add 2 and 3."}]
+    gen_req = await _e2e_render_chat(harmony_client, HARMONY_MODEL, messages)
+
+    reasoning_text = "The user wants 2 plus 3."
+    answer_text = "The answer is 5."
+    assistant_msg = {
+        "role": "assistant",
+        "thinking": reasoning_text,
+        "content": answer_text,
+    }
+    output_ids = _harmony_extract_assistant_ids(harmony_tokenizer, assistant_msg)
+
+    decoded = harmony_tokenizer.decode(output_ids)
+    if reasoning_text not in decoded:
+        pytest.skip("Harmony template did not render thinking")
+
+    resp = await harmony_client.post(
+        "/v1/chat/completions/derender",
+        json={
+            "model": HARMONY_MODEL,
+            "generate_response": _e2e_generate_response(output_ids),
+            "prompt_tokens": len(gen_req["token_ids"]),
+            "chat_request": {
+                "model": HARMONY_MODEL,
+                "messages": messages,
+                "include_reasoning": True,
+            },
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    msg = resp.json()["choices"][0]["message"]
+    assert msg["reasoning"] is not None
+    assert reasoning_text in msg["reasoning"]
+    assert answer_text in (msg["content"] or "")

tests/kernels/quantization/test_per_token_group_quant.py

@@ -345,6 +345,63 @@ def test_per_token_group_quant_fp8_packed_zero_fills_padded_output_q(
         )
 
 
+@pytest.mark.skipif(
+    not current_platform.is_cuda_alike(),
+    reason="packed FP8 per-token-group quant kernel requires a CUDA-alike GPU",
+)
+def test_per_token_group_quant_fp8_packed_large_mn():
+    """Regression test for https://github.com/vllm-project/vllm/issues/45099.
+
+    Some background: gridDim.x and gridDim.y have different limits of 2^31 - 1 and
+    2^16 - 1, respectively.
+    Prior code introduced a bug where it incorrectly assumed grid.x and y both have
+    2^31 - 1 limits and mixed them up, which doesn't surface until the kernel is
+    launched with a large mn that exceeds grid.y limit (2^16 - 1).
+
+    This issue doesn't surface often because each forward pass only processes a
+    bounded token batch, not the full context.
+    Quantizing tensors with more rows than that will fail at launch with
+    "CUDA error: invalid argument".
+    This is a differential test that compares fp8 output against Triton output
+    reference when token size sits just above the gridDim.y 2^16 - 1 limit.
+    """
+
+    device = "cuda"
+    group_size = 128
+    # hidden 2048 -> 2048/128 = 16 groups per row -> kx=16, ry=1: one grid row per mn
+    # row, so any mn > 65535 overflowed grid.y before the fix.
+    num_tokens, hidden_dim = 65537, 2048
+    torch.manual_seed(42)
+    x = torch.randn((num_tokens, hidden_dim), device=device, dtype=torch.bfloat16) * 8
+
+    out_q, out_s_packed = fp8_utils.per_token_group_quant_fp8_packed_for_deepgemm(
+        x,
+        group_size=group_size,
+        use_ue8m0=True,
+    )
+
+    with patch("vllm.platforms.current_platform.is_cuda_alike", return_value=False):
+        ref_q, ref_s = fp8_utils.per_token_group_quant_fp8(
+            x, group_size, use_ue8m0=True
+        )
+
+    assert torch.equal(out_q, ref_q), "Quantized output mismatch"
+
+    # Vectorized packed-scale check; the per-element loop used by the smaller
+    # tests is too slow at this size. groups_per_row is a multiple of 4 here,
+    # so there is no K padding and the packed view lines up.
+    mn = num_tokens
+    groups_per_row = hidden_dim // group_size
+    k_num_packed = (groups_per_row + 3) // 4
+    assert groups_per_row % 4 == 0
+    ref_exponents = (ref_s.reshape(mn, groups_per_row).view(torch.int32) >> 23) & 0xFF
+    exp = ref_exponents.view(mn, k_num_packed, 4)
+    expected = (
+        exp[..., 0] | (exp[..., 1] << 8) | (exp[..., 2] << 16) | (exp[..., 3] << 24)
+    )
+    assert torch.equal(out_s_packed.cpu(), expected.cpu()), "Packed scale mismatch"
+
+
 @pytest.mark.parametrize("shape", [(32, 128), (64, 256), (16, 512)])
 @pytest.mark.parametrize("group_size", [64, 128])
 @pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")

tests/models/multimodal/processing/test_qwen3_vl.py

@@ -92,3 +92,49 @@ def test_processor_num_frames_timestamp(
     assert len(video_phs) == 1, (
         f"Expected exactly 1 video placeholder, got {len(video_phs)}"
     )
+
+
+@pytest.mark.parametrize("model_id", [MODEL_ID])
+@pytest.mark.parametrize("num_videos", [2, 4])
+def test_processor_multi_video(
+    model_id: str,
+    num_videos: int,
+) -> None:
+    """Verify that multi-video processing produces correct placeholders.
+
+    This exercises the token-level replacement path in
+    ``_call_hf_processor`` which avoids the quadratic text-level
+    prompt expansion.
+    """
+    ctx = build_model_context(
+        model_id,
+        limit_mm_per_prompt={"image": 0, "video": num_videos},
+    )
+    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
+
+    prompt = "<|vision_start|><|video_pad|><|vision_end|>" * num_videos
+    mm_data = {"video": [_build_video_mm_data(num_frames=8)["video"][0]] * num_videos}
+
+    processed = processor(
+        prompt,
+        mm_items=processor.info.parse_mm_data(mm_data),
+        hf_processor_mm_kwargs={"num_frames": 8},
+    )
+
+    token_ids = processed["prompt_token_ids"]
+    assert len(token_ids) > 0
+
+    video_phs = processed["mm_placeholders"].get("video", [])
+    assert len(video_phs) == num_videos, (
+        f"Expected {num_videos} video placeholders, got {len(video_phs)}"
+    )
+
+    # All placeholders should have the same length (same video params)
+    # and must not overlap.
+    lengths = {ph.length for ph in video_phs}
+    assert len(lengths) == 1, f"Placeholder lengths differ: {lengths}"
+    for i in range(1, len(video_phs)):
+        prev_end = video_phs[i - 1].offset + video_phs[i - 1].length
+        assert video_phs[i].offset >= prev_end, (
+            f"Placeholder {i} overlaps with placeholder {i - 1}"
+        )

tests/reasoning/test_minimax_m3_reasoning_parser.py

@@ -83,6 +83,20 @@ class MiniMaxM3Tokenizer:
         return "".join(tokens)
 
 
+class SplitMiniMaxM3Tokenizer(MiniMaxM3Tokenizer):
+    """Tokenizer that exposes marker vocab entries but encodes them as text."""
+
+    def tokenize(self, text: str) -> list[str]:
+        return list(text)
+
+
+class RuntimeSplitMiniMaxM3Tokenizer(MiniMaxM3Tokenizer):
+    """Tokenizer whose runtime output splits markers despite atomic encodes."""
+
+    def encode_runtime(self, text: str) -> list[int]:
+        return [self._add_token(token) for token in list(text)]
+
+
 def make_parser(
     chat_template_kwargs: dict[str, str] | None = None,
 ) -> tuple[MiniMaxM3ReasoningParser, MiniMaxM3Tokenizer]:
@@ -105,7 +119,8 @@ def run_streaming(
     reasoning_end_states: list[bool] = []
 
     for chunk in chunks:
-        delta_token_ids = tokenizer.encode(chunk, add_special_tokens=False)
+        encode_runtime = getattr(tokenizer, "encode_runtime", tokenizer.encode)
+        delta_token_ids = encode_runtime(chunk)
         current_text = previous_text + chunk
         current_token_ids = previous_token_ids + delta_token_ids
         delta = parser.extract_reasoning_streaming(
@@ -174,14 +189,14 @@ def test_nonstreaming_drops_leading_end_tag():
     assert content == "answer"
 
 
-def test_nonstreaming_non_leading_end_tag_is_content():
+def test_nonstreaming_end_tag_in_content_state_is_dropped():
     parser, _ = make_parser()
     request = ChatCompletionRequest(messages=[], model="test-model")
 
     reasoning, content = parser.extract_reasoning("XXX</mm:think>YYY", request)
 
     assert reasoning is None
-    assert content == "XXX</mm:think>YYY"
+    assert content == "XXXYYY"
 
 
 def test_nonstreaming_enabled_mode_starts_in_reasoning():
@@ -246,7 +261,7 @@ def test_streaming_drops_leading_end_tag():
     assert end_states == [True, True]
 
 
-def test_streaming_non_leading_end_tag_is_content():
+def test_streaming_end_tag_in_content_state_is_dropped():
     parser, tokenizer = make_parser()
 
     reasoning, content, end_states = run_streaming(
@@ -256,7 +271,7 @@ def test_streaming_non_leading_end_tag_is_content():
     )
 
     assert reasoning is None
-    assert content == "XXX</mm:think>YYY"
+    assert content == "XXXYYY"
     assert end_states == [True]
 
 
@@ -288,6 +303,110 @@ def test_streaming_plain_content_ends_reasoning_phase():
     assert end_states == [True, True]
 
 
+def test_streaming_split_marker_tokens_are_not_returned():
+    tokenizer = RuntimeSplitMiniMaxM3Tokenizer()
+    parser = MiniMaxM3ReasoningParser(tokenizer)
+
+    reasoning, content, end_states = run_streaming(
+        parser,
+        tokenizer,
+        ["<mm:think>", "Reasoning", " content", "</mm:think>", "content"],
+    )
+
+    assert reasoning == "Reasoning content"
+    assert content == "content"
+    assert end_states == [False, False, False, True, True]
+
+
+def test_streaming_split_marker_text_drives_end_state():
+    tokenizer = RuntimeSplitMiniMaxM3Tokenizer()
+    parser = MiniMaxM3ReasoningParser(tokenizer)
+    previous_text = ""
+    previous_token_ids: list[int] = []
+
+    for chunk in ["<mm:think>", "Reasoning", " content", "</mm:think>"]:
+        delta_token_ids = tokenizer.encode_runtime(chunk)
+        current_text = previous_text + chunk
+        current_token_ids = previous_token_ids + delta_token_ids
+        parser.extract_reasoning_streaming(
+            previous_text=previous_text,
+            current_text=current_text,
+            delta_text=chunk,
+            previous_token_ids=previous_token_ids,
+            current_token_ids=current_token_ids,
+            delta_token_ids=delta_token_ids,
+        )
+        previous_text = current_text
+        previous_token_ids = current_token_ids
+
+    assert parser.is_reasoning_end_streaming(previous_token_ids, []) is True
+
+
+def test_streaming_split_marker_tokens_enabled_mode():
+    tokenizer = RuntimeSplitMiniMaxM3Tokenizer()
+    parser = MiniMaxM3ReasoningParser(
+        tokenizer, chat_template_kwargs={"thinking_mode": "enabled"}
+    )
+
+    reasoning, content, end_states = run_streaming(
+        parser,
+        tokenizer,
+        ["Reasoning", " content", "</mm:think>", "content"],
+    )
+
+    assert reasoning == "Reasoning content"
+    assert content == "content"
+    assert end_states == [False, False, True, True]
+
+
+def test_streaming_split_marker_text_across_deltas():
+    tokenizer = RuntimeSplitMiniMaxM3Tokenizer()
+    parser = MiniMaxM3ReasoningParser(tokenizer)
+
+    reasoning, content, end_states = run_streaming(
+        parser,
+        tokenizer,
+        ["<mm:", "think>", "Reasoning", " content", "</mm:", "think>", "content"],
+    )
+
+    assert reasoning == "Reasoning content"
+    assert content == "content"
+    assert end_states == [False, False, False, False, False, True, True]
+
+
+def test_streaming_split_leading_end_marker_text_across_deltas():
+    tokenizer = RuntimeSplitMiniMaxM3Tokenizer()
+    parser = MiniMaxM3ReasoningParser(tokenizer)
+
+    reasoning, content, end_states = run_streaming(
+        parser,
+        tokenizer,
+        ["</mm:", "think>", "content"],
+    )
+
+    assert reasoning is None
+    assert content == "content"
+    assert end_states == [False, True, True]
+
+
+def test_token_id_helpers_with_split_marker_tokens():
+    tokenizer = SplitMiniMaxM3Tokenizer()
+    parser = MiniMaxM3ReasoningParser(tokenizer)
+    output_ids = tokenizer.encode(
+        "<mm:think>abc</mm:think>def", add_special_tokens=False
+    )
+    open_reasoning_ids = tokenizer.encode("<mm:think>abc", add_special_tokens=False)
+    content_ids = tokenizer.encode("plain", add_special_tokens=False)
+
+    assert parser.is_reasoning_end(output_ids)
+    assert not parser.is_reasoning_end(open_reasoning_ids)
+    assert not parser.is_reasoning_end(content_ids)
+    assert tokenizer.decode(parser.extract_content_ids(output_ids)) == "def"
+    assert parser.extract_content_ids(open_reasoning_ids) == []
+    assert parser.extract_content_ids(content_ids) == content_ids
+    assert parser.count_reasoning_tokens(output_ids) == len(tokenizer.encode("abc"))
+
+
 def test_token_id_helpers():
     parser, tokenizer = make_parser()
     output_ids = tokenizer.encode(

tests/v1/core/test_contiguous_kv_packing.py

@@ -1,16 +1,23 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Tests for contiguous KV cache packing in _get_kv_cache_config_deepseek_v4."""
+"""Tests for contiguous KV cache packing."""
 
 from unittest.mock import MagicMock
 
 import pytest
 import torch
 
-from vllm.v1.core.kv_cache_utils import _get_kv_cache_config_deepseek_v4
+from vllm import envs
+from vllm.v1.core.kv_cache_utils import (
+    _get_kv_cache_config_deepseek_v4,
+    get_kv_cache_config_from_groups,
+)
 from vllm.v1.kv_cache_interface import (
+    FullAttentionSpec,
     KVCacheGroupSpec,
+    KVCacheTensor,
     MLAAttentionSpec,
+    SlidingWindowSpec,
     UniformTypeKVCacheSpecs,
 )
 
@@ -28,6 +35,25 @@ def _make_mla_spec(page_size: int, block_size: int = 256) -> MLAAttentionSpec:
     )
 
 
+def _make_full_spec() -> FullAttentionSpec:
+    return FullAttentionSpec(
+        block_size=16,
+        num_kv_heads=2,
+        head_size=64,
+        dtype=torch.float16,
+    )
+
+
+def _make_sw_spec() -> SlidingWindowSpec:
+    return SlidingWindowSpec(
+        block_size=16,
+        num_kv_heads=2,
+        head_size=64,
+        dtype=torch.float16,
+        sliding_window=128,
+    )
+
+
 def _make_groups(n_c4, n_c128, n_swa):
     PS_C4_MLA = 37440
     PS_C4_IDX = 8640
@@ -130,6 +156,73 @@ class TestInterleavedPacking:
         for i, v in enumerate(views):
             assert (v == i + 1).all(), f"View {i} was corrupted"
 
+    def test_hma_attention_groups_keep_default_backing(self, monkeypatch):
+        monkeypatch.setattr(envs, "VLLM_USE_PACKED_HMA_KV_CACHE", False, raising=False)
+        full = _make_full_spec()
+        sw = _make_sw_spec()
+        page_size = full.page_size_bytes
+        groups = [
+            KVCacheGroupSpec(["full.0", "full.1"], full),
+            KVCacheGroupSpec(["sw.0", "sw.2"], sw),
+            KVCacheGroupSpec(["sw.1", "sw.3"], sw),
+        ]
+
+        config = get_kv_cache_config_from_groups(
+            _mock_vllm_config(), groups, available_memory=page_size * 2 * 32
+        )
+
+        assert config.num_blocks == 32
+        assert sum(t.size for t in config.kv_cache_tensors) == page_size * 2 * 32
+        assert config.kv_cache_tensors == [
+            KVCacheTensor(size=page_size * 32, shared_by=["full.0", "sw.0", "sw.1"]),
+            KVCacheTensor(size=page_size * 32, shared_by=["full.1", "sw.2", "sw.3"]),
+        ]
+
+    def test_hma_attention_groups_use_packed_backing_with_flag(self, monkeypatch):
+        monkeypatch.setattr(envs, "VLLM_USE_PACKED_HMA_KV_CACHE", True, raising=False)
+        full = _make_full_spec()
+        sw = _make_sw_spec()
+        page_size = full.page_size_bytes
+        groups = [
+            KVCacheGroupSpec(["full.0", "full.1"], full),
+            KVCacheGroupSpec(["sw.0", "sw.2"], sw),
+            KVCacheGroupSpec(["sw.1", "sw.3"], sw),
+        ]
+
+        config = get_kv_cache_config_from_groups(
+            _mock_vllm_config(), groups, available_memory=page_size * 2 * 32
+        )
+
+        assert config.num_blocks == 32
+        assert {t.size for t in config.kv_cache_tensors} == {page_size * 2 * 32}
+        assert config.kv_cache_tensors == [
+            KVCacheTensor(
+                size=page_size * 2 * 32,
+                shared_by=["full.0", "sw.0", "sw.1"],
+                offset=0,
+                block_stride=page_size * 2,
+            ),
+            KVCacheTensor(
+                size=page_size * 2 * 32,
+                shared_by=["full.1", "sw.2", "sw.3"],
+                offset=page_size,
+                block_stride=page_size * 2,
+            ),
+        ]
+
+    def test_single_group_attention_keeps_unpacked_layout(self):
+        spec = _make_full_spec()
+        groups = [KVCacheGroupSpec(["full.0", "full.1"], spec)]
+
+        config = get_kv_cache_config_from_groups(
+            _mock_vllm_config(), groups, available_memory=spec.page_size_bytes * 2 * 32
+        )
+
+        assert sum(t.size for t in config.kv_cache_tensors) == (
+            spec.page_size_bytes * 2 * 32
+        )
+        assert [t.block_stride for t in config.kv_cache_tensors] == [0, 0]
+
 
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])

tests/v1/core/test_scheduler.py

@@ -144,6 +144,43 @@ def test_async_scheduling_pp_allows_rescheduling_with_output_placeholders():
     assert req.request_id in output.num_scheduled_tokens
 
 
+def test_cached_request_data_resumed_all_token_ids_mrv1_only():
+    """all_token_ids carries a resumed request's token ids to the connector
+    for the V1 model runner, but is skipped entirely for the V2 model runner.
+    """
+    from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+
+    scheduler = create_scheduler()
+    (req,) = create_requests(num_requests=1, num_tokens=8)
+    req.append_output_token_ids([101, 102, 103])
+
+    # A resumed request was not scheduled in the previous step.
+    assert req.request_id not in scheduler.prev_step_scheduled_req_ids
+
+    empty_blocks = KVCacheBlocks(blocks=((),))
+
+    def make_cached():
+        return scheduler._make_cached_request_data(
+            running_reqs=[],
+            resumed_reqs=[req],
+            num_scheduled_tokens={req.request_id: 1},
+            spec_decode_tokens={},
+            req_to_new_blocks={req.request_id: empty_blocks},
+        )
+
+    # V1 model runner: the full token id list is propagated.
+    assert not scheduler.use_v2_model_runner
+    cached = make_cached()
+    assert req.request_id in cached.resumed_req_ids
+    assert cached.all_token_ids[req.request_id] == list(req.all_token_ids)
+
+    # V2 model runner: all_token_ids is skipped entirely.
+    scheduler.use_v2_model_runner = True
+    cached = make_cached()
+    assert req.request_id in cached.resumed_req_ids
+    assert cached.all_token_ids == {}
+
+
 def test_schedule_partial_requests():
     """Test scheduling behavior with partial requests.
 

tests/v1/e2e/general/test_cascade_attention.py

@@ -4,9 +4,16 @@
 import pytest
 
 from vllm import LLM, SamplingParams
+from vllm.platforms import current_platform
 
 from ....utils import create_new_process_for_each_test
 
+if current_platform.is_rocm():
+    pytest.skip(
+        "Cascade attention backends FLASH_ATTN and FLASHINFER are notsupported on ROCm",
+        allow_module_level=True,
+    )
+
 
 @create_new_process_for_each_test()
 @pytest.mark.parametrize("attn_backend", ["FLASH_ATTN", "FLASHINFER"])

tests/v1/kv_connector/unit/test_mooncake_store_coordinator.py

@@ -7,7 +7,10 @@ from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.coordinator imp
     ExternalCachedBlockPool,
     MooncakeStoreCoordinator,
 )
-from vllm.v1.core.kv_cache_utils import BlockHash, BlockHashListWithBlockSize
+from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.data import (
+    chunk_hashes_for_block_size,
+)
+from vllm.v1.core.kv_cache_utils import BlockHash
 from vllm.v1.kv_cache_interface import (
     FullAttentionSpec,
     KVCacheGroupSpec,
@@ -182,7 +185,7 @@ def test_coordinator_group_block_size_double_hash():
     ]
     coord = _make_coord(groups, hash_block_size=16)
     hs = _hashes(4)
-    big_hashes = list(BlockHashListWithBlockSize(hs, 16, 32))
+    big_hashes = list(chunk_hashes_for_block_size(hs, 16, 32))
     exists = {(0, bytes(h)) for h in hs}
     exists |= {(1, bytes(bh)) for bh in big_hashes}
     cmap = ExternalCachedBlockPool(exists)

tests/v1/kv_connector/unit/test_mooncake_store_hma_e2e.py

@@ -323,8 +323,8 @@ def test_recv_skips_swa_blocks_before_window():
 
 def test_chunked_token_database_hash_block_size_smaller_than_block_size():
     """DSv4-style: hash_block_size=4, group block_size=16 — process_tokens
-    must merge every 4 fine hashes into one chunk hash via
-    BlockHashListWithBlockSize."""
+    keys each 16-token chunk by its last fine hash, keeping the Mooncake key
+    at one digest instead of concatenating all 4 fine hashes."""
     md = KeyMetadata("m", 0, 0, 0, 0, group_id=3)
     db = ChunkedTokenDatabase(md, block_size=16, hash_block_size=4)
     db.set_kv_caches_base_addr([0])
@@ -335,8 +335,7 @@ def test_chunked_token_database_hash_block_size_smaller_than_block_size():
     assert len(out) == 2
     assert out[0][0] == 0 and out[0][1] == 16
     assert out[1][0] == 16 and out[1][1] == 32
-    # Each chunk's hash is the concatenation of 4 fine hashes.
-    expected0 = b"".join(fine_hashes[0:4]).hex()
-    expected1 = b"".join(fine_hashes[4:8]).hex()
-    assert out[0][2].chunk_hash == expected0
-    assert out[1][2].chunk_hash == expected1
+    # Each chunk's hash is its last (4th) fine hash, which already chains the
+    # prior three.
+    assert out[0][2].chunk_hash == fine_hashes[3].hex()
+    assert out[1][2].chunk_hash == fine_hashes[7].hex()

tests/v1/kv_connector/unit/test_mooncake_store_worker.py

@@ -23,6 +23,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store import (
     worker as mooncake_store_worker,
 )
 from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.data import (
+    BlobBlockHashes,
     ChunkedTokenDatabase,
     KeyMetadata,
     LoadSpec,
@@ -32,6 +33,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.data import (
 from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.metrics import (
     MooncakeStoreConnectorStats,
 )
+from vllm.v1.core.kv_cache_utils import BlockHash
 
 
 def _default_send_coord() -> mooncake_store_worker.MooncakeStoreCoordinator:
@@ -1179,9 +1181,9 @@ def test_store_sending_thread_kv_events_use_group_chunk_metadata():
     assert full_event.group_idx == 0
     assert full_event.block_size == 32
     assert full_event.token_ids == list(range(32))
-    assert full_event.block_hashes == [
-        maybe_convert_block_hash(BlockHash(b"".join(hs)))
-    ]
+    # block_size=32 over hash_block_size=8 (scale 4): the chunk is keyed by its
+    # last sub-hash, not the concatenation of all four.
+    assert full_event.block_hashes == [maybe_convert_block_hash(BlockHash(hs[3]))]
 
     assert swa_event.group_idx == 1
     assert swa_event.block_size == 8
@@ -1749,3 +1751,33 @@ def test_store_worker_close_swallows_store_errors():
     worker.close()
 
     assert worker.store is None
+
+
+def test_blob_block_hashes_wire_roundtrip():
+    """The lookup wire format sends a ``hash_len`` frame plus the raw hashes
+    concatenated back-to-back; the server rebuilds them through a zero-copy
+    ``BlobBlockHashes`` view over the frame buffer."""
+    hashes = [BlockHash(bytes([i]) * 16) for i in range(5)]
+    hash_len = len(hashes[0])
+
+    # Client side (LookupKeyClient._lookup): flat payload frame.
+    blob = b"".join(hashes)
+
+    # Server side (LookupKeyServer): view over the frame buffer (a memoryview),
+    # never materializing the full hash list upfront.
+    view = BlobBlockHashes(memoryview(blob), hash_len)
+
+    assert len(view) == 5
+    assert list(view) == hashes  # default Sequence iter terminates via IndexError
+    assert [bytes(h) for h in view] == hashes
+    assert bytes(view[-1]) == hashes[-1]
+    assert [bytes(h) for h in view[1:3]] == hashes[1:3]
+    with pytest.raises(IndexError):
+        _ = view[5]
+
+
+def test_blob_block_hashes_empty():
+    """Empty lookups send hash_len=0 and an empty payload."""
+    view = BlobBlockHashes(memoryview(b""), 0)
+    assert len(view) == 0
+    assert list(view) == []

tests/v1/kv_offload/cpu/test_manager.py

@@ -14,12 +14,13 @@ from vllm.v1.kv_offload.base import (
     ReqContext,
     make_offload_key,
 )
-from vllm.v1.kv_offload.cpu.common import CPULoadStoreSpec
+from vllm.v1.kv_offload.cpu.common import (
+    CPULoadStoreSpec,
+    CPUOffloadingMetrics,
+)
 from vllm.v1.kv_offload.cpu.manager import CPUOffloadingManager
 from vllm.v1.kv_offload.cpu.policies.arc import ARCCachePolicy
 
-STORES_SKIPPED = "vllm:kv_offload_stores_skipped"
-
 
 def make_req_context(
     req_id: str = "", kv_transfer_params: dict | None = None
@@ -181,10 +182,45 @@ def test_filter_reused_manager_reports_stores_skipped_counter():
     )
     stats = manager.get_stats()
     assert stats is not None
-    assert stats.reduce()[STORES_SKIPPED] == 3
+    assert stats.reduce()[CPUOffloadingMetrics.STORES_SKIPPED] == 3
     stats = manager.get_stats()
     assert stats is not None
-    assert stats.reduce()[STORES_SKIPPED] == 0
+    assert stats.reduce()[CPUOffloadingMetrics.STORES_SKIPPED] == 0
+
+
+def test_cpu_manager_reports_cache_usage_gauge():
+    def check_usage_stats(manager: CPUOffloadingManager, value: float):
+        stats = manager.get_stats()
+        assert stats is not None
+        assert stats.reduce()[
+            CPUOffloadingMetrics.CPU_CACHE_USAGE_PERC
+        ] == pytest.approx(value)
+
+    # Zero-capacity manager always reports 0.0
+    manager = make_cpu_manager(num_blocks=0)
+    check_usage_stats(manager, 0.0)
+
+    # Empty manager (4 blocks, none allocated): usage = 0.0
+    manager = make_cpu_manager(num_blocks=4)
+    check_usage_stats(manager, 0.0)
+
+    # After allocating 2 of 4 blocks: usage = 0.5
+    manager.prepare_store(to_keys([1, 2]), _EMPTY_REQ_CTX)
+    check_usage_stats(manager, 0.5)
+
+    # After filling all 4 blocks: usage = 1.0
+    manager.prepare_store(to_keys([3, 4]), _EMPTY_REQ_CTX)
+    check_usage_stats(manager, 1.0)
+
+    # After completing store, the blocks becomes evictable as it is not actively used
+    # and usage drops.
+    manager.complete_store(to_keys([1, 2]), _EMPTY_REQ_CTX)
+    check_usage_stats(manager, 0.5)
+
+    # After completing store, the blocks becomes evictable as it is not actively used
+    # and usage drops.
+    manager.complete_store(to_keys([3, 4]), _EMPTY_REQ_CTX)
+    check_usage_stats(manager, 0.0)
 
 
 def test_cpu_manager():

tests/v1/logits_processors/test_correctness.py

@@ -145,7 +145,6 @@ def _generate_fake_sampling_metadata(
         vllm_config.scheduler_config.max_num_seqs,
         num_spec,
         device,
-        PIN_MEMORY_AVAILABLE,
     )
     fake_sampling_metadata = SamplingMetadata(
         temperature=torch.full((batch_size,), 0.0),
@@ -880,7 +879,6 @@ def test_maybe_create_thinking_budget_holder_without_reasoning():
             cfg.scheduler_config.max_num_seqs,
             0,
             torch.device("cpu"),
-            False,
         )
         is None
     )

tests/v1/simple_kv_offload/test_scheduler.py

@@ -6,6 +6,7 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 
+import pytest
 import torch
 
 from vllm import SamplingParams
@@ -1528,3 +1529,232 @@ def test_reset_pending_loads() -> None:
     # All GPU blocks free
     num_used = gpu_pool.num_gpu_blocks - gpu_pool.get_num_free_blocks()
     assert num_used == 1, f"Expected only null block in use, got {num_used}"
+
+
+def _make_cp_vllm_config(
+    dcp_world_size: int = 1,
+    pcp_world_size: int = 1,
+) -> VllmConfig:
+    """VllmConfig with context-parallel sizes set for scheduler-only tests."""
+    cfg = _make_vllm_config()
+
+    cfg.parallel_config.decode_context_parallel_size = dcp_world_size
+    cfg.parallel_config.prefill_context_parallel_size = pcp_world_size
+    return cfg
+
+
+def _make_cp_scheduler(
+    *,
+    dcp_world_size: int = 1,
+    pcp_world_size: int = 1,
+    num_cpu_blocks: int = 8,
+    num_gpu_blocks: int = 16,
+    lazy: bool = False,
+) -> SchedulerFixture:
+    """Build a SimpleCPUOffloadScheduler with CP-scaled virtual block size."""
+    cp_world_size = dcp_world_size * pcp_world_size
+    virtual_block_size = BLOCK_SIZE * cp_world_size
+
+    kv_cache_config = _make_kv_cache_config(num_gpu_blocks)
+    vllm_config = _make_cp_vllm_config(dcp_world_size, pcp_world_size)
+    cpu_capacity_bytes = _BYTES_PER_BLOCK * num_cpu_blocks
+
+    sched = SimpleCPUOffloadScheduler(
+        vllm_config=vllm_config,
+        kv_cache_config=kv_cache_config,
+        cpu_capacity_bytes=cpu_capacity_bytes,
+        scheduler_block_size=virtual_block_size,
+        hash_block_size=virtual_block_size,
+        lazy_offload=lazy,
+    )
+
+    gpu_block_pool = BlockPool(
+        num_gpu_blocks=num_gpu_blocks,
+        enable_caching=True,
+        hash_block_size=virtual_block_size,
+    )
+    sched.bind_gpu_block_pool(gpu_block_pool)
+
+    return SchedulerFixture(
+        scheduler=sched,
+        gpu_block_pool=gpu_block_pool,
+        vllm_config=vllm_config,
+        kv_cache_config=kv_cache_config,
+    )
+
+
+def _make_cp_request(
+    num_blocks: int,
+    virtual_block_size: int,
+    request_id: str | None = None,
+) -> Request:
+    """Create a request whose block hashes are computed at the virtual
+    (CP-scaled) block size, matching what the real scheduler does.
+    """
+    global _req_counter
+    _req_counter += 1
+    if request_id is None:
+        request_id = f"req-cp-{_req_counter}"
+
+    num_tokens = num_blocks * virtual_block_size + 1
+    start = _req_counter * 10000
+    prompt_token_ids = list(range(start, start + num_tokens))
+    sampling_params = SamplingParams(max_tokens=1)
+
+    return Request(
+        request_id=request_id,
+        prompt_token_ids=prompt_token_ids,
+        sampling_params=sampling_params,
+        pooling_params=None,
+        mm_features=None,
+        block_hasher=get_request_block_hasher(virtual_block_size, sha256),
+    )
+
+
+def _allocate_cp_gpu_blocks(
+    gpu_block_pool: BlockPool,
+    request: Request,
+    num_blocks: int,
+    virtual_block_size: int,
+    group_id: int = 0,
+) -> list:
+    """Allocate GPU blocks and cache them using the CP-scaled block size."""
+    blocks = gpu_block_pool.get_new_blocks(num_blocks)
+    num_full = min(num_blocks, len(request.block_hashes))
+    if num_full > 0:
+        gpu_block_pool.cache_full_blocks(
+            request=request,
+            blocks=blocks,
+            num_cached_blocks=0,
+            num_full_blocks=num_full,
+            block_size=virtual_block_size,
+            kv_cache_group_id=group_id,
+        )
+    return blocks
+
+
+# ---------------------------------------------------------------------------
+# Test 15: CP block size scaling is correct
+# ---------------------------------------------------------------------------
+@pytest.mark.parametrize(
+    "dcp_world_size, pcp_world_size",
+    [
+        (2, 1),  # DCP only
+        (1, 2),  # PCP only
+        (2, 2),  # DCP + PCP
+    ],
+)
+def test_cp_block_size_scaling(dcp_world_size: int, pcp_world_size: int) -> None:
+    """Verify that the scheduler's block_size and cp_world_size are correctly
+    scaled when context parallelism is enabled."""
+    fix = _make_cp_scheduler(
+        dcp_world_size=dcp_world_size, pcp_world_size=pcp_world_size
+    )
+    sched = fix.scheduler
+
+    expected_cp = dcp_world_size * pcp_world_size
+    assert sched.cp_world_size == expected_cp
+    assert sched.block_size == BLOCK_SIZE * expected_cp
+
+
+# ---------------------------------------------------------------------------
+# Test 16: CP eager store-and-load roundtrip
+# ---------------------------------------------------------------------------
+@pytest.mark.parametrize(
+    "dcp_world_size, pcp_world_size",
+    [
+        (2, 1),
+        (1, 2),
+    ],
+)
+def test_cp_eager_store_and_load_roundtrip(
+    dcp_world_size: int, pcp_world_size: int
+) -> None:
+    """With CP enabled, store blocks to CPU and reload them for a new request
+    with matching tokens.  Verifies that hash matching and transfer-pair
+    construction work with the virtual block size."""
+    fix = _make_cp_scheduler(
+        dcp_world_size=dcp_world_size,
+        pcp_world_size=pcp_world_size,
+        num_cpu_blocks=8,
+        num_gpu_blocks=16,
+        lazy=False,
+    )
+    sched = fix.scheduler
+    cp = dcp_world_size * pcp_world_size
+    vbs = BLOCK_SIZE * cp
+
+    num_blocks = 2
+    req = _make_cp_request(num_blocks, vbs)
+
+    # Allocate GPU blocks and register hashes
+    gpu_blocks = _allocate_cp_gpu_blocks(fix.gpu_block_pool, req, num_blocks, vbs)
+    kv_blocks = KVCacheBlocks(blocks=(gpu_blocks,))
+    req.num_computed_tokens = num_blocks * vbs
+    sched.update_state_after_alloc(req, kv_blocks, num_external_tokens=0)
+
+    block_ids = kv_blocks.get_block_ids()
+    sched_out = make_scheduler_output(
+        {req.request_id: num_blocks * vbs},
+        new_reqs={req.request_id: block_ids},
+    )
+
+    meta = sched.build_connector_meta(sched_out)
+    assert meta.store_event >= 0, "Expected a store event"
+    assert len(meta.store_gpu_blocks) == num_blocks
+    assert len(meta.store_cpu_blocks) == num_blocks
+    simulate_store_completion(sched, meta.store_event)
+
+    # New request with same tokens — should get a full CPU cache hit.
+    req2 = Request(
+        request_id="req-cp-load",
+        prompt_token_ids=req.prompt_token_ids,
+        sampling_params=req.sampling_params,
+        pooling_params=None,
+        mm_features=None,
+        block_hasher=req._block_hasher,
+    )
+
+    hit_tokens, is_async = sched.get_num_new_matched_tokens(req2, num_computed_tokens=0)
+    assert hit_tokens == num_blocks * vbs
+    assert is_async is True
+
+    # Allocate fresh GPU blocks for the load.
+    gpu_blocks2 = fix.gpu_block_pool.get_new_blocks(num_blocks)
+    kv_blocks2 = KVCacheBlocks(blocks=(gpu_blocks2,))
+    sched.update_state_after_alloc(req2, kv_blocks2, num_external_tokens=hit_tokens)
+
+    sched_out2 = make_scheduler_output(
+        {req2.request_id: 1},
+        new_reqs={req2.request_id: kv_blocks2.get_block_ids()},
+    )
+    meta2 = sched.build_connector_meta(sched_out2)
+    assert meta2.load_event >= 0, "Expected a load event"
+    assert len(meta2.load_gpu_blocks) == num_blocks
+    assert len(meta2.load_cpu_blocks) == num_blocks
+
+
+# ---------------------------------------------------------------------------
+# Test 17: CP lazy target blocks are scaled correctly
+# ---------------------------------------------------------------------------
+@pytest.mark.parametrize("cp_world_size", [1, 2, 4])
+def test_cp_lazy_target_blocks_scaling(cp_world_size: int) -> None:
+    """_estimate_lazy_target_blocks returns fewer blocks when cp_world_size > 1
+    because each virtual block covers more tokens."""
+    kv_cache_config = _make_kv_cache_config(num_blocks=16)
+    max_batched = 64
+
+    target_base = SimpleCPUOffloadScheduler._estimate_lazy_target_blocks(
+        kv_cache_config, max_batched, cp_world_size=1
+    )
+    target_cp = SimpleCPUOffloadScheduler._estimate_lazy_target_blocks(
+        kv_cache_config, max_batched, cp_world_size=cp_world_size
+    )
+
+    if cp_world_size == 1:
+        assert target_cp == target_base
+    else:
+        assert target_cp < target_base, (
+            f"cp_world_size={cp_world_size}: target_cp={target_cp} should be "
+            f"less than target_base={target_base}"
+        )

tests/v1/streaming_input/test_gpu_model_runner_streaming.py

@@ -35,7 +35,6 @@ def mock_model_runner_with_input_batch():
         max_model_len=1024,
         max_num_batched_tokens=1024,
         device="cpu",
-        pin_memory=False,
         vocab_size=32000,
         block_sizes=[16],
         kernel_block_sizes=[16],

tests/v1/worker/test_gpu_input_batch.py

@@ -10,7 +10,6 @@ import torch
 
 from vllm.platforms import current_platform
 from vllm.sampling_params import SamplingParams
-from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.torch_utils import make_tensor_with_pad
 from vllm.v1.pool.metadata import PoolingMetadata
 from vllm.v1.sample.logits_processor import LogitsProcessors
@@ -236,7 +235,6 @@ def test_sampling_metadata_in_input_batch(device: str, batch_size: int):
         max_model_len=1024,
         max_num_batched_tokens=1024,
         device=torch.device(device),
-        pin_memory=is_pin_memory_available(),
         vocab_size=1024,
         block_sizes=[1],
         kernel_block_sizes=[1],
@@ -331,7 +329,6 @@ def test_swap_states_in_input_batch(device: str, batch_size: int, swap_list: lis
         max_model_len=1024,
         max_num_batched_tokens=1024,
         device=torch.device(device),
-        pin_memory=is_pin_memory_available(),
         vocab_size=1024,
         block_sizes=[1],
         kernel_block_sizes=[1],
@@ -341,7 +338,6 @@ def test_swap_states_in_input_batch(device: str, batch_size: int, swap_list: lis
         max_model_len=1024,
         max_num_batched_tokens=1024,
         device=torch.device(device),
-        pin_memory=is_pin_memory_available(),
         vocab_size=1024,
         block_sizes=[1],
         kernel_block_sizes=[1],
@@ -410,7 +406,6 @@ def test_pooling_prompt_lens_not_aliased(device: str):
         max_model_len=MAX_PROMPT_SIZE + NUM_OUTPUT_TOKENS,
         max_num_batched_tokens=batch_size * (MAX_PROMPT_SIZE + NUM_OUTPUT_TOKENS),
         device=torch.device(device),
-        pin_memory=is_pin_memory_available(),
         vocab_size=VOCAB_SIZE,
         block_sizes=[16],
         kernel_block_sizes=[16],
@@ -459,7 +454,6 @@ def test_pooling_metadata_token_id_buffers(
         max_model_len=MAX_PROMPT_SIZE + NUM_OUTPUT_TOKENS,
         max_num_batched_tokens=MAX_PROMPT_SIZE + NUM_OUTPUT_TOKENS,
         device=torch.device("cpu"),
-        pin_memory=False,
         vocab_size=VOCAB_SIZE,
         block_sizes=[16],
         kernel_block_sizes=[16],

tests/v1/worker/test_gpu_model_runner.py

@@ -85,7 +85,6 @@ def initialize_kv_cache(runner: GPUModelRunner):
         max_model_len=runner.max_model_len,
         max_num_batched_tokens=runner.max_num_tokens,
         device=runner.device,
-        pin_memory=runner.pin_memory,
         vocab_size=runner.model_config.get_vocab_size(),
         block_sizes=[kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size],
         kernel_block_sizes=[
@@ -1405,7 +1404,6 @@ def test_input_batch_with_kernel_block_sizes():
     max_model_len = 512
     max_num_batched_tokens = 512
     device = torch.device(DEVICE_TYPE)
-    pin_memory = False
     vocab_size = 50272
 
     # Test with different kernel block sizes
@@ -1417,7 +1415,6 @@ def test_input_batch_with_kernel_block_sizes():
         max_model_len=max_model_len,
         max_num_batched_tokens=max_num_batched_tokens,
         device=device,
-        pin_memory=pin_memory,
         vocab_size=vocab_size,
         block_sizes=block_sizes,
         kernel_block_sizes=kernel_block_sizes,
@@ -1478,7 +1475,6 @@ def test_hybrid_cache_integration(default_vllm_config, dist_init):
         max_model_len=runner.max_model_len,
         max_num_batched_tokens=runner.max_num_tokens,
         device=runner.device,
-        pin_memory=runner.pin_memory,
         vocab_size=runner.model_config.get_vocab_size(),
         block_sizes=[kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size],
         kernel_block_sizes=[16],

vllm/compilation/backends.py

@@ -991,7 +991,7 @@ class VllmBackend:
             },
             payload_fn=lambda: json.dumps(
                 {
-                    "model": self.vllm_config.model_config.model,
+                    "model": getattr(self.vllm_config.model_config, "model", "unknown"),
                     "prefix": self.prefix,
                     "mode": str(cc.mode),
                     "backend": cc.backend,

vllm/config/parallel.py

@@ -302,6 +302,14 @@ class ParallelConfig:
     Each entry must use `numactl --physcpubind` CPU-list syntax, for example
     `"0-3"` or `"0,2,4-7"`.
     """
+    assigned_physical_gpu_ids: list[int] | None = None
+    """Mapping from vLLM-local logical GPU IDs to physical GPU IDs.
+
+    For example, ``[2, 3]`` means logical GPU 0 maps to physical GPU 2,
+    and logical GPU 1 maps to physical GPU 3. Physical IDs are used only
+    at platform/topology boundaries such as NVML, NIC affinity, P2P
+    checks, and final CUDA device selection when needed. When None,
+    logical IDs map to visible device IDs in order."""
 
     distributed_timeout_seconds: int | None = None
     """Timeout in seconds for distributed operations (e.g., init_process_group).
@@ -772,6 +780,7 @@ class ParallelConfig:
             "numa_bind",
             "numa_bind_nodes",
             "numa_bind_cpus",
+            "assigned_physical_gpu_ids",
         }
 
         from vllm.config.utils import get_hash_factors, hash_factors

vllm/device_allocator/cumem.py

@@ -18,8 +18,8 @@ import torch
 
 from vllm.device_allocator import AllocationData, HandleType
 from vllm.logger import init_logger
-from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.system_utils import find_loaded_library
+from vllm.utils.torch_utils import PIN_MEMORY
 
 logger = init_logger(__name__)
 
@@ -196,7 +196,7 @@ class CuMemAllocator:
                     size_in_bytes,
                     dtype=torch.uint8,
                     device="cpu",
-                    pin_memory=is_pin_memory_available(),
+                    pin_memory=PIN_MEMORY,
                 )
                 cpu_ptr = cpu_backup_tensor.data_ptr()
                 libcudart.cudaMemcpy(cpu_ptr, ptr, size_in_bytes)

vllm/device_allocator/xpumem.py

@@ -11,7 +11,7 @@ import torch
 
 from vllm.device_allocator import AllocationData, HandleType
 from vllm.logger import init_logger
-from vllm.utils.platform_utils import is_pin_memory_available
+from vllm.utils.torch_utils import PIN_MEMORY
 
 logger = init_logger(__name__)
 
@@ -188,7 +188,7 @@ class XpuMemAllocator:
                 size_in_bytes,
                 dtype=torch.uint8,
                 device="cpu",
-                pin_memory=is_pin_memory_available(),
+                pin_memory=PIN_MEMORY,
             )
             cpu_ptr = cpu_backup_tensor.data_ptr()
             _xpu_memcpy_sync(

vllm/distributed/device_communicators/all2all.py

@@ -704,7 +704,14 @@ class FlashInferNVLinkOneSidedManager(All2AllManagerBase):
         self.num_experts = num_experts
 
         self.cleanup()
-        gpus_per_node = torch.accelerator.device_count()
+        from vllm.platforms.interface import get_assigned_physical_gpu_ids
+
+        assigned_physical_gpu_ids = get_assigned_physical_gpu_ids()
+        gpus_per_node = (
+            len(assigned_physical_gpu_ids)
+            if assigned_physical_gpu_ids is not None
+            else torch.accelerator.device_count()
+        )
         logger.debug(
             "Making One-sided NVLink mapping: rank=%d, world size=%d",
             self.rank,

vllm/distributed/device_communicators/all_reduce_utils.py

@@ -320,13 +320,21 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
 
     is_distributed = dist.is_initialized()
 
-    num_dev = current_platform.device_count()
-    cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
-    if cuda_visible_devices is None:
-        cuda_visible_devices = ",".join(str(i) for i in range(num_dev))
+    from vllm.platforms.interface import get_assigned_physical_gpu_ids
+
+    assigned_physical_gpu_ids = get_assigned_physical_gpu_ids()
+    if assigned_physical_gpu_ids is not None:
+        # Key by the ordered list: the cache stores directed local-index
+        # pairs, so permutations of the same set are distinct mappings.
+        cache_key = ",".join(str(i) for i in assigned_physical_gpu_ids)
+        num_dev = len(assigned_physical_gpu_ids)
+    else:
+        num_dev = current_platform.device_count()
+        cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
+        cache_key = cuda_visible_devices or ",".join(str(i) for i in range(num_dev))
 
     path = os.path.join(
-        envs.VLLM_CACHE_ROOT, f"gpu_p2p_access_cache_for_{cuda_visible_devices}.json"
+        envs.VLLM_CACHE_ROOT, f"gpu_p2p_access_cache_for_{cache_key}.json"
     )
     os.makedirs(os.path.dirname(path), exist_ok=True)
     from vllm.distributed.parallel_state import get_world_group
@@ -338,7 +346,15 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
         #  enter this block to calculate the cache
         logger.info("generating GPU P2P access cache in %s", path)
         cache: dict[str, bool] = {}
-        ids = list(range(num_dev))
+        # The probe subprocesses inherit this process's device-control env
+        # var, so they must be given visible ordinals, not physical IDs.
+        if assigned_physical_gpu_ids is not None:
+            ids = [
+                current_platform.logical_device_id_to_visible_device_id(local)
+                for local in range(num_dev)
+            ]
+        else:
+            ids = list(range(num_dev))
         # batch of all pairs of GPUs
         batch_src, batch_tgt = zip(*list(product(ids, ids)))
         # NOTE: we use `subprocess` rather than `multiprocessing` here
@@ -368,8 +384,11 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
                 ) from e
             with open(output_file.name, "rb") as f:
                 result = pickle.load(f)
+        # Cache entries must be keyed by local indices (0..N-1) because
+        # gpu_p2p_access_check() is called with local ranks.
+        id_to_local = {device_id: local for local, device_id in enumerate(ids)}
         for _i, _j, r in zip(batch_src, batch_tgt, result):
-            cache[f"{_i}->{_j}"] = r
+            cache[f"{id_to_local[_i]}->{id_to_local[_j]}"] = r
         with open(path, "w") as f:
             json.dump(cache, f, indent=4)
     if is_distributed:

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -34,7 +34,12 @@ def _can_p2p(rank: int, world_size: int) -> bool:
             continue
         if envs.VLLM_SKIP_P2P_CHECK:
             logger.debug("Skipping P2P check and trusting the driver's P2P report.")
-            return torch.cuda.can_device_access_peer(rank, i)
+            # can_device_access_peer takes visible device ordinals, while
+            # rank and i are logical local IDs.
+            return torch.cuda.can_device_access_peer(
+                current_platform.logical_device_id_to_visible_device_id(rank),
+                current_platform.logical_device_id_to_visible_device_id(i),
+            )
         if not gpu_p2p_access_check(rank, i):
             return False
     return True
@@ -126,13 +131,10 @@ class CustomAllreduce:
                     CUSTOM_ALL_REDUCE_MAX_SIZES[device_capability_str][world_size],
                     max_size,
                 )
-        cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
-        if cuda_visible_devices:
-            device_ids = list(map(int, cuda_visible_devices.split(",")))
-        else:
-            device_ids = list(range(current_platform.device_count()))
-
-        physical_device_id = device_ids[device.index]
+        # device.index is a visible ordinal, not a logical local ID.
+        physical_device_id = current_platform.visible_device_id_to_physical_device_id(
+            device.index
+        )
         tensor = torch.tensor([physical_device_id], dtype=torch.int, device="cpu")
         gather_list = [
             torch.tensor([0], dtype=torch.int, device="cpu") for _ in range(world_size)

vllm/distributed/device_communicators/quick_all_reduce.py

@@ -129,12 +129,10 @@ class QuickAllReduce:
         assert isinstance(device, torch.device)
         self.device = device
 
-        cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
-        if cuda_visible_devices:
-            device_ids = list(map(int, cuda_visible_devices.split(",")))
-        else:
-            device_ids = list(range(current_platform.device_count()))
-        physical_device_id = device_ids[device.index]
+        # device.index is a visible ordinal, not a logical local ID.
+        physical_device_id = current_platform.visible_device_id_to_physical_device_id(
+            device.index
+        )
         tensor = torch.tensor([physical_device_id], dtype=torch.int, device="cpu")
         gather_list = [
             torch.tensor([0], dtype=torch.int, device="cpu")

vllm/distributed/device_communicators/shm_broadcast.py

@@ -840,7 +840,13 @@ class MessageQueue:
             The MessageQueue instance for the calling process,
             and a list of handles (only non-empty for the reader process).
         """
-        local_size = current_platform.device_count()
+        from vllm.platforms.interface import get_assigned_physical_gpu_ids
+
+        assigned_physical_gpu_ids = get_assigned_physical_gpu_ids()
+        if assigned_physical_gpu_ids is not None:
+            local_size = len(assigned_physical_gpu_ids)
+        else:
+            local_size = current_platform.device_count()
         rank = dist.get_rank()
         same_node = rank // local_size == reader_rank // local_size
         buffer_io = MessageQueue(

vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/vllm_v1_adapter.py

@@ -482,10 +482,11 @@ def _init_lmcache_engine(
     )
 
     # Change current device.
-    num_gpus = torch.accelerator.device_count()
-    local_rank = parallel_config.rank % num_gpus
-    torch.accelerator.set_device_index(local_rank)
-    device = torch.device(f"cuda:{local_rank}")
+    from vllm.distributed.parallel_state import get_world_group
+
+    device_index = get_world_group().device_index
+    torch.accelerator.set_device_index(device_index)
+    device = torch.device(f"cuda:{device_index}")
     metadata = LMCacheEngineMetadata(
         model_config.model,
         parallel_config.world_size,

vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/coordinator.py

@@ -2,13 +2,15 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """External-store cache-hit coordinator for MooncakeStoreConnector."""
 
+from collections.abc import Sequence
 from typing import cast
 
+from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.data import (
+    chunk_hashes_for_block_size,
+)
 from vllm.v1.core.block_pool import BlockPool
 from vllm.v1.core.kv_cache_utils import (
     BlockHash,
-    BlockHashList,
-    BlockHashListWithBlockSize,
     KVCacheBlock,
 )
 from vllm.v1.core.single_type_kv_cache_manager import (
@@ -120,7 +122,7 @@ class MooncakeStoreCoordinator:
 
     def find_longest_cache_hit(
         self,
-        block_hashes: list[BlockHash],
+        block_hashes: Sequence[BlockHash],
         max_length: int,
         cached_block_pool: ExternalCachedBlockPool,
         *,
@@ -147,7 +149,7 @@ class MooncakeStoreCoordinator:
 
     def load_mask(
         self,
-        block_hashes: list[BlockHash],
+        block_hashes: Sequence[BlockHash],
         token_len: int,
     ) -> tuple[list[bool], ...]:
         """Per-group load masks: ``mask[g][i]`` is True iff group ``g``'s
@@ -236,17 +238,15 @@ class MooncakeStoreCoordinator:
         return tuple(masks)
 
     def block_hashes_for_spec(
-        self, block_hashes: list[BlockHash], spec: KVCacheSpec
-    ) -> BlockHashList:
-        if spec.block_size == self.hash_block_size:
-            return block_hashes
-        return BlockHashListWithBlockSize(
+        self, block_hashes: Sequence[BlockHash], spec: KVCacheSpec
+    ) -> Sequence[BlockHash]:
+        return chunk_hashes_for_block_size(
             block_hashes, self.hash_block_size, spec.block_size
         )
 
     def _find_hit_blocks(
         self,
-        block_hashes: list[BlockHash],
+        block_hashes: Sequence[BlockHash],
         max_length: int,
         cached_block_pool: ExternalCachedBlockPool,
         *,
@@ -264,7 +264,7 @@ class MooncakeStoreCoordinator:
             spec, group_ids, manager_cls = self.attention_groups[0]
             hashes = self.block_hashes_for_spec(block_hashes, spec)
             hit_blocks = manager_cls.find_longest_cache_hit(
-                block_hashes=hashes,
+                block_hashes=hashes,  # type: ignore[arg-type]
                 max_length=max_length,
                 kv_cache_group_ids=group_ids,
                 block_pool=cast(BlockPool, cached_block_pool),
@@ -304,7 +304,7 @@ class MooncakeStoreCoordinator:
                     _max_length = min(curr_hit_length + spec.block_size, max_length)
                 hashes = self.block_hashes_for_spec(block_hashes, spec)
                 hit_blocks = manager_cls.find_longest_cache_hit(
-                    block_hashes=hashes,
+                    block_hashes=hashes,  # type: ignore[arg-type]
                     max_length=_max_length,
                     kv_cache_group_ids=group_ids,
                     block_pool=cast(BlockPool, cached_block_pool),

vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/data.py

@@ -5,8 +5,9 @@
 # (vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/).
 """Data classes for MooncakeStoreConnector."""
 
-from collections.abc import Iterable
+from collections.abc import Iterable, Sequence
 from dataclasses import dataclass
+from typing import cast
 
 import torch
 
@@ -23,6 +24,77 @@ from vllm.v1.core.kv_cache_utils import (
 logger = init_logger(__name__)
 
 
+class BlobBlockHashes(Sequence[BlockHash]):
+    """Lazy view over a flat buffer of fixed-size block hashes to avoid the overhead
+    of materializing all hashes upfront.
+    """
+
+    def __init__(self, blob: memoryview, hash_len: int):
+        self._blob = blob
+        self._hash_len = hash_len
+        self._n = len(blob) // hash_len if hash_len else 0
+
+    def __len__(self) -> int:
+        return self._n
+
+    def __getitem__(self, idx):
+        if isinstance(idx, slice):
+            return [self[i] for i in range(*idx.indices(self._n))]
+        if idx < 0:
+            idx += self._n
+        if not 0 <= idx < self._n:
+            raise IndexError(idx)
+        off = idx * self._hash_len
+        return BlockHash(self._blob[off : off + self._hash_len])
+
+
+class _CompactChunkHashList(BlockHashListWithBlockSize):
+    """View that keys each ``block_size`` chunk by the last constituent
+    ``hash_block_size`` hash instead of concatenating all of them.
+
+    The engine chains block hashes (each hash folds in the previous one), so the
+    final sub-block hash of a chunk already uniquely identifies the whole chunk
+    and its prefix. Using it keeps a Mooncake key at a single hash digest
+    regardless of the ``block_size`` / ``hash_block_size`` ratio, instead of
+    growing the key linearly with it (e.g. 64x for ``block_size=256``,
+    ``hash_block_size=4``).
+    """
+
+    def __init__(
+        self,
+        block_hashes: Sequence[BlockHash],
+        hash_block_size: int,
+        target_block_size: int,
+    ):
+        # Accept any indexable sequence (e.g. the lazy ``BlobBlockHashes``), not
+        # just ``list``; the base only indexes/sizes it.
+        assert target_block_size % hash_block_size == 0
+        self.block_hashes = block_hashes  # type: ignore[assignment]
+        self.scale_factor = target_block_size // hash_block_size
+
+    def _get_value_at(self, idx: int) -> BlockHash:
+        return self.block_hashes[idx * self.scale_factor + self.scale_factor - 1]
+
+
+def chunk_hashes_for_block_size(
+    block_hashes: Sequence[BlockHash],
+    hash_block_size: int,
+    block_size: int,
+) -> Sequence[BlockHash]:
+    """Map ``hash_block_size``-granular block hashes to one compact hash per
+    ``block_size`` chunk (the chunk's last sub-hash). Returns ``block_hashes``
+    unchanged when the two sizes are equal.
+    """
+    if block_size == hash_block_size:
+        return block_hashes
+    # Structurally a Sequence[BlockHash] (indexable + sized); the base class
+    # just isn't declared as one.
+    return cast(
+        "Sequence[BlockHash]",
+        _CompactChunkHashList(block_hashes, hash_block_size, block_size),
+    )
+
+
 @dataclass
 class KeyMetadata:
     """Metadata for constructing pool keys."""
@@ -138,18 +210,15 @@ class ChunkedTokenDatabase:
         Args:
             token_len: Total number of tokens.
             block_hashes: Block hashes computed at ``hash_block_size`` granularity.
-                When ``block_size > hash_block_size`` consecutive hashes are merged
-                up to the group's ``block_size`` via ``BlockHashListWithBlockSize``.
+                When ``block_size > hash_block_size`` each group's ``block_size`` chunk
+                is keyed by its last sub-hash via ``chunk_hashes_for_block_size``.
             mask_num: Number of tokens to skip from the beginning.
         """
         if not block_hashes:
             return
-        if self.block_size == self.hash_block_size:
-            chunk_hashes: Iterable[BlockHash] = block_hashes
-        else:
-            chunk_hashes = BlockHashListWithBlockSize(
-                block_hashes, self.hash_block_size, self.block_size
-            )
+        chunk_hashes: Iterable[BlockHash] = chunk_hashes_for_block_size(
+            block_hashes, self.hash_block_size, self.block_size
+        )
         for chunk_id, h in enumerate(chunk_hashes):
             start_idx = chunk_id * self.block_size
             if start_idx >= token_len:

vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/protocol.py

@@ -11,7 +11,10 @@ Wire format (REQ/REP over IPC):
 
       msg_type == LOOKUP_MSG:
           frame 1: token_len (u32 big-endian, 4 bytes)
-          frame 2..n: msgpack-encoded list[str] of block-hash hex digests
+          frame 2: hash_len (u16 big-endian, 2 bytes) — byte length of each
+                   fixed-size block hash (0 when there are no hashes)
+          frame 3: raw block hashes concatenated back-to-back (each hash_len
+                   bytes); the server splits on hash_len
         Response: [hit_count: u32 big-endian, 4 bytes]
 
       msg_type == RESET_MSG:

vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/worker.py

@@ -18,7 +18,7 @@ import socket
 import threading
 import time
 from collections import defaultdict
-from collections.abc import Callable
+from collections.abc import Callable, Sequence
 from concurrent.futures import Future, ThreadPoolExecutor
 from dataclasses import dataclass
 from typing import Any, Literal, TypeVar
@@ -45,6 +45,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.coordinator imp
     MooncakeStoreCoordinator,
 )
 from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.data import (  # noqa: E501
+    BlobBlockHashes,
     ChunkedTokenDatabase,
     KeyMetadata,
     MooncakeStoreConnectorMetadata,
@@ -65,7 +66,6 @@ from vllm.v1.core.kv_cache_utils import (
     resolve_kv_cache_block_sizes,
 )
 from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheGroupSpec
-from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder
 
 from .metrics import MooncakeStoreConnectorStats
 
@@ -1372,7 +1372,7 @@ class MooncakeStoreWorker:
 
         return finished_sending
 
-    def lookup(self, token_len: int, block_hashes: list[BlockHash]) -> int:
+    def lookup(self, token_len: int, block_hashes: Sequence[BlockHash]) -> int:
         """Check how many prefix tokens exist in the store.
 
         Checks across all TP ranks and PP ranks.
@@ -1392,6 +1392,11 @@ class MooncakeStoreWorker:
             group_hashes = self.coord.block_hashes_for_spec(
                 block_hashes, self._kv_cache_groups[g_idx].kv_cache_spec
             )
+            metadata_templates = [
+                dataclasses.replace(db.metadata, tp_rank=tp, pp_rank=pp)
+                for tp in range(tp_count)
+                for pp in range(self.pp_size)
+            ]
             for chunk_id, h in enumerate(group_hashes):
                 start_idx = chunk_id * spec_block_size
                 if start_idx >= token_len:
@@ -1400,11 +1405,11 @@ class MooncakeStoreWorker:
                     chunk_id >= len(lookup_mask) or not lookup_mask[chunk_id]
                 ):
                     continue
-                for tp in range(tp_count):
-                    for pp in range(self.pp_size):
-                        md = dataclasses.replace(db.metadata, tp_rank=tp, pp_rank=pp)
-                        candidate_keys.append(PoolKey(md, h.hex()).to_string())
-                        candidate_meta.append((g_idx, bytes(h)))
+                h_hex = h.hex()
+                h_bytes = bytes(h)
+                for md in metadata_templates:
+                    candidate_keys.append(PoolKey(md, h_hex).to_string())
+                    candidate_meta.append((g_idx, h_bytes))
 
         if not candidate_keys:
             return 0
@@ -1483,7 +1488,6 @@ class LookupKeyServer:
         store_worker: MooncakeStoreWorker,
         vllm_config: VllmConfig,
     ):
-        self.decoder = MsgpackDecoder()
         self.ctx = zmq.Context()  # type: ignore[attr-defined]
         socket_path = get_zmq_rpc_path_lookup(vllm_config)
         self._ipc_path = socket_path.removeprefix("ipc://")
@@ -1506,9 +1510,9 @@ class LookupKeyServer:
 
                 if msg_type == LOOKUP_MSG:
                     token_len = int.from_bytes(all_frames[1], byteorder="big")
-                    hash_frames = all_frames[2:]
-                    hashes_str = self.decoder.decode(hash_frames)
-                    block_hashes = [BlockHash(bytes.fromhex(s)) for s in hashes_str]
+                    hash_len = int.from_bytes(all_frames[2], byteorder="big")
+                    blob = all_frames[3].buffer
+                    block_hashes = BlobBlockHashes(blob, hash_len)
                     result = self.store_worker.lookup(token_len, block_hashes)
                     self.socket.send(result.to_bytes(4, "big"))
 
@@ -1557,7 +1561,6 @@ class LookupKeyClient:
     """
 
     def __init__(self, vllm_config: VllmConfig):
-        self.encoder = MsgpackEncoder()
         self.ctx = zmq.Context()  # type: ignore[attr-defined]
         socket_path = get_zmq_rpc_path_lookup(vllm_config)
         self.socket = make_zmq_socket(
@@ -1574,14 +1577,16 @@ class LookupKeyClient:
         self.futures: dict[str, Future[int]] = {}
 
     def _lookup(self, token_len: int, block_hashes: list[BlockHash]) -> int:
-        hash_strs = [h.hex() for h in block_hashes]
-        hash_frames = self.encoder.encode(hash_strs)
-        token_len_bytes = token_len.to_bytes(4, byteorder="big")
-        all_frames = [LOOKUP_MSG, token_len_bytes] + list(hash_frames)
+        hash_len = len(block_hashes[0]) if block_hashes else 0
+        all_frames = (
+            LOOKUP_MSG,
+            token_len.to_bytes(4, byteorder="big"),
+            hash_len.to_bytes(2, byteorder="big"),
+            b"".join(block_hashes),
+        )
         self.socket.send_multipart(all_frames, copy=False)
         resp = self.socket.recv()
-        result = int.from_bytes(resp, "big")
-        return result
+        return int.from_bytes(resp, "big")
 
     def lookup(
         self,

vllm/distributed/kv_transfer/kv_connector/v1/offloading/worker.py

@@ -50,7 +50,8 @@ class OffloadingConnectorWorker:
     def register_kv_caches(
         self, kv_caches: dict[str, torch.Tensor | list[torch.Tensor]]
     ):
-        num_blocks = self.spec.kv_cache_config.num_blocks
+        kv_cache_config = self.spec.kv_cache_config
+        num_blocks = kv_cache_config.num_blocks
 
         # layer_name -> (num_blocks, page_size_bytes) tensor
         tensors_per_block: dict[str, tuple[torch.Tensor, ...]] = {}
@@ -58,7 +59,7 @@ class OffloadingConnectorWorker:
         unpadded_page_size_bytes: dict[str, int] = {}
         # layer_name -> size of page in bytes
         page_size_bytes: dict[str, int] = {}
-        for kv_cache_group in self.spec.kv_cache_config.kv_cache_groups:
+        for kv_cache_group in kv_cache_config.kv_cache_groups:
             group_layer_names = kv_cache_group.layer_names
             group_kv_cache_spec = kv_cache_group.kv_cache_spec
             if isinstance(group_kv_cache_spec, UniformTypeKVCacheSpecs):
@@ -122,9 +123,35 @@ class OffloadingConnectorWorker:
                 else:
                     raise NotImplementedError
 
+        packed_kv_cache_tensor = next(
+            (t for t in kv_cache_config.kv_cache_tensors if t.block_stride), None
+        )
+        is_dsv4 = all(
+            isinstance(group.kv_cache_spec, UniformTypeKVCacheSpecs)
+            for group in kv_cache_config.kv_cache_groups
+        )
+        if packed_kv_cache_tensor is not None and not is_dsv4:
+            (tensor,) = tensors_per_block[packed_kv_cache_tensor.shared_by[0]]
+            block_stride = tensor.stride(0)
+            packed_tensor = tensor.as_strided(
+                (num_blocks, block_stride),
+                (block_stride, 1),
+                storage_offset=0,
+            )
+            self._register_handlers(
+                CanonicalKVCaches(
+                    [CanonicalKVCacheTensor(packed_tensor, block_stride)],
+                    [
+                        [CanonicalKVCacheRef(0, block_stride)]
+                        for _ in kv_cache_config.kv_cache_groups
+                    ],
+                )
+            )
+            return
+
         block_tensors: list[CanonicalKVCacheTensor] = []
         block_data_refs: dict[str, list[CanonicalKVCacheRef]] = defaultdict(list)
-        for kv_cache_tensor in self.spec.kv_cache_config.kv_cache_tensors:
+        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
             # Filter to layers that were actually processed above.
             # _get_kv_cache_config_deepseek_v4 emits KVCacheTensor entries for
             # every (tuple_idx, page_size) slot; slots where no group has a
@@ -166,7 +193,7 @@ class OffloadingConnectorWorker:
                     )
 
         group_data_refs: list[list[CanonicalKVCacheRef]] = []
-        for kv_cache_group in self.spec.kv_cache_config.kv_cache_groups:
+        for kv_cache_group in kv_cache_config.kv_cache_groups:
             group_refs: list[CanonicalKVCacheRef] = []
             for layer_name in kv_cache_group.layer_names:
                 group_refs += block_data_refs[layer_name]

vllm/distributed/parallel_state.py

@@ -392,6 +392,14 @@ class GroupCoordinator:
 
         self.rank = torch.distributed.get_rank()
         self.local_rank = local_rank
+        self.device_index: int
+        if _WORLD is not None:
+            self.device_index = _WORLD.device_index
+        else:
+            assert local_rank >= 0, (
+                "local_rank must be provided when creating the world group"
+            )
+            self.device_index = local_rank
 
         self_device_group = None
         self_cpu_group = None
@@ -442,11 +450,18 @@ class GroupCoordinator:
         from vllm.platforms import current_platform
 
         if current_platform.is_cuda_alike():
-            self.device = torch.device(f"cuda:{local_rank}")
+            visible_device_index = (
+                current_platform.logical_device_id_to_visible_device_id(
+                    self.device_index
+                )
+            )
+            self.device = torch.device(f"cuda:{visible_device_index}")
         elif current_platform.is_xpu():
-            self.device = torch.device(f"xpu:{local_rank}")
+            self.device = torch.device(f"xpu:{self.device_index}")
         elif current_platform.is_out_of_tree():
-            self.device = torch.device(f"{current_platform.device_name}:{local_rank}")
+            self.device = torch.device(
+                f"{current_platform.device_name}:{self.device_index}"
+            )
         else:
             self.device = torch.device("cpu")
 
@@ -1438,7 +1453,12 @@ def _init_process_group_for_split_group(
     """
     if torch.accelerator.is_available() and backend != "gloo":
         init_backend = "cpu:gloo,cuda:nccl"
-        device_id: torch.device | None = torch.device(f"cuda:{local_rank}")
+        from vllm.platforms import current_platform
+
+        visible_device_index = current_platform.logical_device_id_to_visible_device_id(
+            local_rank
+        )
+        device_id: torch.device | None = torch.device(f"cuda:{visible_device_index}")
     else:
         init_backend = "gloo"
         device_id = None

vllm/distributed/stateless_coordinator.py

@@ -86,6 +86,15 @@ class StatelessGroupCoordinator(GroupCoordinator):
 
         self.rank = global_rank
         self.local_rank = local_rank
+        from vllm.distributed.parallel_state import _WORLD
+
+        if _WORLD is not None:
+            self.device_index = _WORLD.device_index
+        else:
+            assert local_rank >= 0, (
+                "local_rank must be provided when creating the world group"
+            )
+            self.device_index = local_rank
 
         self_device_group = None
         self_cpu_group = None
@@ -152,11 +161,18 @@ class StatelessGroupCoordinator(GroupCoordinator):
         self.tcp_store_group = self_tcp_store_group
 
         if current_platform.is_cuda_alike():
-            self.device = torch.device(f"cuda:{local_rank}")
+            visible_device_index = (
+                current_platform.logical_device_id_to_visible_device_id(
+                    self.device_index
+                )
+            )
+            self.device = torch.device(f"cuda:{visible_device_index}")
         elif current_platform.is_xpu():
-            self.device = torch.device(f"xpu:{local_rank}")
+            self.device = torch.device(f"xpu:{self.device_index}")
         elif current_platform.is_out_of_tree():
-            self.device = torch.device(f"{current_platform.device_name}:{local_rank}")
+            self.device = torch.device(
+                f"{current_platform.device_name}:{self.device_index}"
+            )
         else:
             self.device = torch.device("cpu")
 

vllm/engine/arg_utils.py

@@ -6,6 +6,7 @@ import copy
 import dataclasses
 import functools
 import json
+import os
 import sys
 from collections.abc import Callable
 from dataclasses import MISSING, asdict, dataclass, fields, is_dataclass
@@ -465,6 +466,7 @@ class EngineArgs:
     numa_bind: bool = ParallelConfig.numa_bind
     numa_bind_nodes: list[int] | None = ParallelConfig.numa_bind_nodes
     numa_bind_cpus: list[str] | None = ParallelConfig.numa_bind_cpus
+    device_ids: list[int | str] | None = None
     tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
     prefill_context_parallel_size: int = ParallelConfig.prefill_context_parallel_size
     decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
@@ -979,6 +981,20 @@ class EngineArgs:
         parallel_group.add_argument(
             "--numa-bind-cpus", **parallel_kwargs["numa_bind_cpus"]
         )
+        parallel_group.add_argument(
+            "--device-ids",
+            type=lambda s: [
+                int(device_id) if device_id.isdigit() else device_id
+                for device_id in (part.strip() for part in s.split(","))
+            ],
+            default=None,
+            help="Comma-separated physical GPU device IDs or UUIDs to use "
+            '(e.g. --device-ids "2,3,5,7"). Avoids setting '
+            "CUDA_VISIBLE_DEVICES, preserving full GPU topology "
+            "visibility for GPU-NIC affinity and DeepGEMM. "
+            "Note: has no effect with Ray executors; use Ray "
+            "placement groups for GPU selection instead.",
+        )
         parallel_group.add_argument(
             "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
         )
@@ -1716,6 +1732,47 @@ class EngineArgs:
         )
         return SpeculativeConfig(**self.speculative_config)
 
+    def _resolve_device_ids(self) -> list[int] | None:
+        if not self.device_ids:
+            return None
+        if self.distributed_executor_backend == "ray":
+            logger.warning(
+                "--device-ids has no effect when using the Ray executor. "
+                "Use Ray placement groups for GPU selection instead."
+            )
+        ids = self.device_ids
+        if len(set(ids)) != len(ids):
+            raise ValueError(f"--device-ids must not contain duplicates: {ids}")
+        if all(isinstance(i, str) for i in ids):
+            return [
+                current_platform.device_control_id_to_physical_device_id(i)
+                for i in cast(list[str], ids)
+            ]
+        if any(isinstance(i, str) for i in ids):
+            raise ValueError("--device-ids must not mix integer IDs and UUIDs")
+        int_ids = cast(list[int], ids)
+        # Compose with CUDA_VISIBLE_DEVICES: if CVD is set, treat
+        # --device-ids values as indices into the CVD-visible set.
+        cvd = getattr(
+            envs,
+            current_platform.device_control_env_var,
+            os.environ.get(current_platform.device_control_env_var),
+        )
+        if cvd:
+            cvd_ids = [
+                current_platform.device_control_id_to_physical_device_id(x)
+                for x in cvd.split(",")
+            ]
+            for i in int_ids:
+                if i >= len(cvd_ids):
+                    raise ValueError(
+                        f"--device-ids index {i} is out of range for "
+                        f"{current_platform.device_control_env_var}"
+                        f"={cvd} ({len(cvd_ids)} devices visible)"
+                    )
+            return [cvd_ids[i] for i in int_ids]
+        return int_ids
+
     def create_diffusion_config(self) -> DiffusionConfig | None:
         if self.diffusion_config is None:
             return None
@@ -2029,6 +2086,7 @@ class EngineArgs:
             cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size,
             _api_process_count=self._api_process_count,
             _api_process_rank=self._api_process_rank,
+            assigned_physical_gpu_ids=self._resolve_device_ids(),
             numa_bind=self.numa_bind,
             numa_bind_nodes=self.numa_bind_nodes,
             numa_bind_cpus=self.numa_bind_cpus,

vllm/entrypoints/anthropic/serving.py

@@ -43,6 +43,7 @@ from vllm.entrypoints.openai.engine.protocol import (
     JsonSchemaResponseFormat,
     ResponseFormat,
     StreamOptions,
+    UsageInfo,
 )
 from vllm.entrypoints.openai.models.serving import OpenAIServingModels
 from vllm.entrypoints.serve.utils.api_utils import sanitize_message
@@ -54,6 +55,49 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+def _get_cached_tokens(usage: UsageInfo | None) -> int | None:
+    """Extract cached token count from OpenAI UsageInfo."""
+    if usage is None or usage.prompt_tokens_details is None:
+        return None
+    return usage.prompt_tokens_details.cached_tokens
+
+
+def _build_anthropic_usage(
+    prompt_tokens: int,
+    completion_tokens: int | None,
+    usage: UsageInfo | None,
+) -> AnthropicUsage:
+    """Build an AnthropicUsage from OpenAI-style token counts.
+
+    Anthropic defines ``total_input == input_tokens + cache_read +
+    cache_creation``.  vLLM's ``prompt_tokens`` is the total, so
+    ``input_tokens = prompt_tokens - cached_tokens``.
+
+    OpenAI usage only exposes ``cached_tokens`` (hits); there is no
+    cache-creation analog, so ``cache_creation_input_tokens`` is ``0``
+    when cache info is present.  When cache info is absent (e.g.
+    ``--enable-prompt-tokens-details`` off, or a streaming chunk that
+    hasn't carried it yet), cache fields are left **unset** so
+    ``exclude_unset=True`` serialization omits them entirely.
+
+    ``completion_tokens`` follows ``UsageInfo`` and may be ``None`` on
+    intermediate stream chunks; we coerce to ``0`` for the wire format.
+    """
+    output_tokens = completion_tokens or 0
+    cached = _get_cached_tokens(usage)
+    if cached is not None:
+        return AnthropicUsage(
+            input_tokens=prompt_tokens - cached,
+            output_tokens=output_tokens,
+            cache_read_input_tokens=cached,
+            cache_creation_input_tokens=0,
+        )
+    return AnthropicUsage(
+        input_tokens=prompt_tokens,
+        output_tokens=output_tokens,
+    )
+
+
 def wrap_data_with_event(data: str, event: str):
     return f"event: {event}\ndata: {data}\n\n"
 
@@ -582,9 +626,10 @@ class AnthropicServingMessages(OpenAIServingChat):
             id=generator.id,
             content=[],
             model=generator.model,
-            usage=AnthropicUsage(
-                input_tokens=generator.usage.prompt_tokens,
-                output_tokens=generator.usage.completion_tokens,
+            usage=_build_anthropic_usage(
+                generator.usage.prompt_tokens,
+                generator.usage.completion_tokens,
+                generator.usage,
             ),
             kv_transfer_params=generator.kv_transfer_params,
         )
@@ -765,11 +810,12 @@ class AnthropicServingMessages(OpenAIServingChat):
                                     model=origin_chunk.model,
                                     stop_reason=None,
                                     stop_sequence=None,
-                                    usage=AnthropicUsage(
-                                        input_tokens=origin_chunk.usage.prompt_tokens
+                                    usage=_build_anthropic_usage(
+                                        origin_chunk.usage.prompt_tokens
                                         if origin_chunk.usage
                                         else 0,
-                                        output_tokens=0,
+                                        0,
+                                        origin_chunk.usage,
                                     ),
                                 ),
                             )
@@ -788,13 +834,14 @@ class AnthropicServingMessages(OpenAIServingChat):
                             chunk = AnthropicStreamEvent(
                                 type="message_delta",
                                 delta=AnthropicDelta(stop_reason=stop_reason),
-                                usage=AnthropicUsage(
-                                    input_tokens=origin_chunk.usage.prompt_tokens
+                                usage=_build_anthropic_usage(
+                                    origin_chunk.usage.prompt_tokens
                                     if origin_chunk.usage
                                     else 0,
-                                    output_tokens=origin_chunk.usage.completion_tokens
+                                    origin_chunk.usage.completion_tokens
                                     if origin_chunk.usage
                                     else 0,
+                                    origin_chunk.usage,
                                 ),
                             )
                             data = chunk.model_dump_json(exclude_unset=True)

vllm/entrypoints/llm.py

@@ -898,12 +898,6 @@ class LLM(BeamSearchOfflineMixin, PoolingOfflineMixin, OfflineInferenceMixin):
     def finish_weight_update(self) -> None:
         """Finish the current weight update."""
         self.llm_engine.collective_rpc("finish_weight_update")
-        # Invalidate cached state computed with the old weights so it isn't
-        # reused for subsequent requests:
-        # - prefix cache: KV blocks computed with the old weights
-        # - encoder cache: multimodal embeddings keyed only by mm_hash
-        self.llm_engine.reset_prefix_cache()
-        self.llm_engine.reset_encoder_cache()
 
     def __repr__(self) -> str:
         """Return a transformers-style hierarchical view of the model."""

vllm/entrypoints/openai/api_server.py

@@ -455,7 +455,7 @@ async def init_render_app_state(
         enable_auto_tools=args.enable_auto_tool_choice,
         exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
         tool_parser=args.tool_call_parser,
-        reasoning_parser=args.structured_outputs_config.reasoning_parser,
+        reasoning_parser=args.reasoning_parser,
         default_chat_template_kwargs=args.default_chat_template_kwargs,
         log_error_stack=args.log_error_stack,
     )

vllm/entrypoints/openai/dp_supervisor.py

@@ -23,12 +23,10 @@ import uvloop
 from fastapi import FastAPI, Response
 
 from vllm.logger import init_logger
-from vllm.platforms import current_platform
 from vllm.utils.system_utils import (
     decorate_logs,
     kill_process_tree,
     set_process_title,
-    update_environment_variables,
 )
 
 logger = init_logger(__name__)
@@ -127,22 +125,29 @@ def _build_vllm_dp_server_args(
     child_args.data_parallel_multi_port_external_lb = False
     child_args.data_parallel_supervisor_port = None
     child_args.api_server_count = 1
+    child_args.device_ids = _build_device_ids(args, local_rank)
     return child_args
 
 
-def _build_vllm_dp_server_env(
-    args: argparse.Namespace, local_rank: int
-) -> dict[str, str]:
-    # set visible devices for the child process
+def _build_device_ids(args: argparse.Namespace, local_rank: int) -> list[int | str]:
+    """Build the --device-ids value for a DP child process.
+
+    The child resolves these against its own inherited device-control env
+    var (e.g. CUDA_VISIBLE_DEVICES), so integer IDs must stay env-relative
+    here rather than being translated to physical IDs.
+    """
     devices_per_rank = args.tensor_parallel_size * args.pipeline_parallel_size
     start = local_rank * devices_per_rank
     stop = start + devices_per_rank
-    device_env = current_platform.device_control_env_var
-    visible_devices = ",".join(
-        str(current_platform.device_id_to_physical_device_id(idx))
-        for idx in range(start, stop)
-    )
-    return {device_env: visible_devices}
+    device_ids = getattr(args, "device_ids", None)
+    if device_ids is not None:
+        if stop > len(device_ids):
+            raise ValueError(
+                f"--device-ids has {len(device_ids)} entries, but DP rank "
+                f"{local_rank} needs devices [{start}, {stop})"
+            )
+        return device_ids[start:stop]
+    return list(range(start, stop))
 
 
 def _child_base_url(args: argparse.Namespace, port: int) -> str:
@@ -228,9 +233,7 @@ def _build_dp_supervisor_app(supervisor: DPSupervisor) -> FastAPI:
     return app
 
 
-def _run_vllm_dp_server(
-    child_args: argparse.Namespace, env_updates: dict[str, str]
-) -> None:
+def _run_vllm_dp_server(child_args: argparse.Namespace) -> None:
     """
     Entrypoint function for the vLLM DP Server.
     """
@@ -241,7 +244,6 @@ def _run_vllm_dp_server(
     os.setpgrp()
 
     name = f"APIServer_DP{child_args.data_parallel_rank}"
-    update_environment_variables(env_updates)
     set_process_title(name)
     decorate_logs(name)
     uvloop.run(run_server(child_args))
@@ -345,11 +347,10 @@ class DPSupervisor:
         context = multiprocessing.get_context("spawn")
         for local_rank in range(self.args.data_parallel_size_local):
             child_args = _build_vllm_dp_server_args(self.args, local_rank)
-            child_env = _build_vllm_dp_server_env(self.args, local_rank)
             process = context.Process(
                 target=_run_vllm_dp_server,
                 name=f"APIServer_DPRank_{child_args.data_parallel_rank}",
-                args=(child_args, child_env),
+                args=(child_args,),
             )
             process.start()
             self._processes.append(process)

vllm/entrypoints/serve/disagg/protocol.py

@@ -219,10 +219,14 @@ class GenerateResponse(BaseModel):
 
 
 class DerenderChatRequest(BaseModel):
-    """Request for the /v1/chat/completions/derender endpoint.
+    """Request for the /v1/chat/completions/derender endpoint (non-streaming).
 
-    Wraps a GenerateResponse and caller-supplied metadata needed to produce
-    a fully-formed ChatCompletionResponse without a GPU.
+    Wraps a complete GenerateResponse and caller-supplied metadata needed to
+    produce a fully-formed ChatCompletionResponse without a GPU.
+
+    Streaming derender would require a separate endpoint design with
+    incremental token delivery, ``OutputProcessor``-based detokenization,
+    and ``parser.parse_delta()`` instead of ``parser.parse()``.
     """
 
     model: str
@@ -244,7 +248,7 @@ class DerenderChatRequest(BaseModel):
 
 
 class DerenderCompletionRequest(BaseModel):
-    """Request for the /v1/completions/derender endpoint.
+    """Request for the /v1/completions/derender endpoint (non-streaming).
 
     Parallel to DerenderChatRequest but handles the multi-prompt completions
     case: one GenerateResponse per prompt, mirroring the list[GenerateRequest]

vllm/entrypoints/serve/render/serving.py

@@ -27,6 +27,7 @@ from vllm.entrypoints.openai.completion.protocol import (
 )
 from vllm.entrypoints.openai.engine.protocol import (
     ErrorResponse,
+    ToolCall,
     UsageInfo,
 )
 from vllm.entrypoints.openai.engine.serving import resolve_token_id_placeholder
@@ -43,7 +44,6 @@ from vllm.entrypoints.serve.disagg.protocol import (
     DerenderChatRequest,
     DerenderCompletionRequest,
     GenerateRequest,
-    GenerateResponseChoice,
     MultiModalFeatures,
     PlaceholderRangeInfo,
 )
@@ -76,21 +76,83 @@ from vllm.utils.mistral import mt as _mt
 logger = init_logger(__name__)
 
 
+def _parse_token_id_placeholder(token: str) -> int | None:
+    """Extract token ID from a 'token_id:N' placeholder string."""
+    if not token.startswith("token_id:"):
+        return None
+    try:
+        return int(token[len("token_id:") :])
+    except ValueError:
+        return None
+
+
+def _correct_decoded_token(
+    token_id: int, context_token_ids: list[int], tokenizer: TokenizerLike
+) -> str:
+    """Use preceding tokens as context to fix U+FFFD from byte-fallback.
+
+    Mirrors LogprobsProcessor._correct_decoded_token in v1/engine/logprobs.py.
+    """
+    max_ctx = min(len(context_token_ids), 4)
+
+    for num_ctx in range(1, max_ctx + 1):
+        context = context_token_ids[-num_ctx:]
+        full_decoded = tokenizer.decode(context + [token_id])
+
+        if full_decoded.endswith("�"):
+            continue
+
+        clean_end = len(context)
+        for j in range(len(context) - 1, -1, -1):
+            if tokenizer.decode([context[j]]).endswith("�"):
+                clean_end = j
+            else:
+                break
+
+        clean_prefix = tokenizer.decode(context[:clean_end]) if clean_end > 0 else ""
+
+        if full_decoded.startswith(clean_prefix):
+            return full_decoded[len(clean_prefix) :]
+
+        common_len = 0
+        for a, b in zip(clean_prefix, full_decoded):
+            if a != b:
+                break
+            common_len += 1
+        return full_decoded[common_len:]
+
+    return ""
+
+
 def _resolve_logprobs(
     logprobs: ChatCompletionLogProbs, tokenizer: TokenizerLike
 ) -> ChatCompletionLogProbs:
-    """Resolve all token_id:N placeholders in a ChatCompletionLogProbs object."""
+    """Resolve token_id:N placeholders in a ChatCompletionLogProbs object."""
     if logprobs.content is None:
         return logprobs
+
+    context_token_ids: list[int] = []
     resolved_content = []
+
     for entry in logprobs.content:
         token_str, token_bytes = resolve_token_id_placeholder(entry.token, tokenizer)
+        sampled_id = _parse_token_id_placeholder(entry.token)
+
+        if token_str.endswith("�") and sampled_id is not None:
+            token_str = _correct_decoded_token(sampled_id, context_token_ids, tokenizer)
+            token_bytes = list(token_str.encode("utf-8"))
+
         resolved_top = []
         for top in entry.top_logprobs:
             top_str, top_bytes = resolve_token_id_placeholder(top.token, tokenizer)
+            top_id = _parse_token_id_placeholder(top.token)
+            if top_str.endswith("�") and top_id is not None:
+                top_str = _correct_decoded_token(top_id, context_token_ids, tokenizer)
+                top_bytes = list(top_str.encode("utf-8"))
             resolved_top.append(
                 top.model_copy(update={"token": top_str, "bytes": top_bytes})
             )
+
         resolved_content.append(
             entry.model_copy(
                 update={
@@ -100,6 +162,10 @@ def _resolve_logprobs(
                 }
             )
         )
+
+        if sampled_id is not None:
+            context_token_ids.append(sampled_id)
+
     return ChatCompletionLogProbs(content=resolved_content)
 
 
@@ -136,30 +202,6 @@ def _convert_chat_logprobs_to_completion_logprobs(
     )
 
 
-def _build_chat_choice(
-    choice: GenerateResponseChoice, tokenizer: TokenizerLike
-) -> ChatCompletionResponseChoice:
-    """Detokenize and resolve logprobs for a single GenerateResponseChoice.
-
-    Raises:
-        ValueError: if choice.token_ids is empty or None.
-    """
-    if not choice.token_ids:
-        raise ValueError(f"choice {choice.index} has empty or null token_ids")
-    decoded_text = tokenizer.decode(choice.token_ids, skip_special_tokens=True)
-    resolved_logprobs = (
-        _resolve_logprobs(choice.logprobs, tokenizer)
-        if choice.logprobs is not None
-        else None
-    )
-    return ChatCompletionResponseChoice(
-        index=choice.index,
-        message=ChatMessage(role="assistant", content=decoded_text),
-        logprobs=resolved_logprobs,
-        finish_reason=choice.finish_reason,
-    )
-
-
 class OpenAIServingRender:
     def __init__(
         self,
@@ -536,9 +578,12 @@ class OpenAIServingRender:
     ) -> ChatCompletionResponse | ErrorResponse:
         """Postprocess a GenerateResponse into a ChatCompletionResponse.
 
-        This is the symmetric inverse of render_chat_request: it detokenizes
-        output token IDs, resolves token_id:N logprob placeholders, and
-        formats the result as an OpenAI-compatible chat completion response.
+        Non-streaming only: expects the complete GenerateResponse with all
+        token IDs present.  Uses ``parser.parse()`` for one-shot extraction.
+
+        When ``request.chat_request`` is provided, the parser splits the
+        output into (reasoning, content, tool_calls).  Otherwise falls
+        back to plain detokenization.
         """
         error_check_ret = await self._check_model(request)
         if error_check_ret is not None:
@@ -546,11 +591,89 @@ class OpenAIServingRender:
 
         tokenizer = self.renderer.get_tokenizer()
         gen = request.generate_response
+        chat_request = request.chat_request
         choices: list[ChatCompletionResponseChoice] = []
 
         try:
             for choice in gen.choices:
-                choices.append(_build_chat_choice(choice, tokenizer))
+                if not choice.token_ids:
+                    raise ValueError(
+                        f"choice {choice.index} has empty or null token_ids"
+                    )
+
+                resolved_logprobs = (
+                    _resolve_logprobs(choice.logprobs, tokenizer)
+                    if choice.logprobs is not None
+                    else None
+                )
+
+                if self.parser is not None and chat_request is not None:
+                    # Parser path: decode with special tokens preserved
+                    # so the parser can see markers like </think>,
+                    # <tool_call>, or Harmony channel tokens.
+                    decoded_text = tokenizer.decode(
+                        choice.token_ids, skip_special_tokens=False
+                    )
+
+                    chat_template_kwargs: dict[str, Any] = {}
+                    if not self.use_harmony:
+                        chat_template_kwargs = (
+                            chat_request.build_chat_params(
+                                self.chat_template,
+                                self.chat_template_content_format,
+                            )
+                            .with_defaults(self.default_chat_template_kwargs)
+                            .chat_template_kwargs
+                        )
+
+                    parser = self.parser(
+                        tokenizer,
+                        chat_request.tools,
+                        chat_template_kwargs=chat_template_kwargs,
+                    )
+                    reasoning, content, tool_calls = parser.parse(
+                        decoded_text,
+                        chat_request,
+                        enable_auto_tools=self.enable_auto_tools,
+                        model_output_token_ids=choice.token_ids,
+                    )
+
+                    if not getattr(chat_request, "include_reasoning", True):
+                        reasoning = None
+
+                    tc_items = (
+                        [
+                            ToolCall(
+                                id=random_uuid(),
+                                function=tc,
+                            )
+                            for tc in tool_calls
+                        ]
+                        if tool_calls
+                        else []
+                    )
+
+                    message = ChatMessage(
+                        role="assistant",
+                        reasoning=reasoning,
+                        content=content,
+                        tool_calls=tc_items,
+                    )
+                else:
+                    # No parser: plain detokenization.
+                    decoded_text = tokenizer.decode(
+                        choice.token_ids, skip_special_tokens=True
+                    )
+                    message = ChatMessage(role="assistant", content=decoded_text)
+
+                choices.append(
+                    ChatCompletionResponseChoice(
+                        index=choice.index,
+                        message=message,
+                        logprobs=resolved_logprobs,
+                        finish_reason=choice.finish_reason,
+                    )
+                )
         except ValueError as exc:
             return self.create_error_response(str(exc))
 
@@ -587,8 +710,9 @@ class OpenAIServingRender:
     ) -> CompletionResponse | ErrorResponse:
         """Postprocess a list of GenerateResponses into a CompletionResponse.
 
-        Mirrors the multi-prompt completions case: one GenerateResponse per
-        prompt, parallel to the list[GenerateRequest] from /v1/completions/render.
+        Non-streaming only.  Mirrors the multi-prompt completions case: one
+        GenerateResponse per prompt, parallel to the list[GenerateRequest]
+        from /v1/completions/render.
         """
         error_check_ret = await self._check_model(request)
         if error_check_ret is not None:

vllm/envs.py

@@ -209,6 +209,7 @@ if TYPE_CHECKING:
     VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
     VLLM_WORKER_SHUTDOWN_TIMEOUT_SECONDS: int = 5
     VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
+    VLLM_USE_PACKED_HMA_KV_CACHE: bool = False
     VLLM_SSM_CONV_STATE_LAYOUT: Literal["SD", "DS"] | None = None
     VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
     VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
@@ -485,7 +486,7 @@ def get_vllm_port() -> int | None:
             raise ValueError(
                 f"VLLM_PORT '{port}' appears to be a URI. "
                 "This may be caused by a Kubernetes service discovery issue,"
-                "check the warning in: https://docs.vllm.ai/en/stable/serving/env_vars.html"
+                "check the warning in: https://docs.vllm.ai/en/latest/configuration/env_vars.html"
             ) from None
         raise ValueError(f"VLLM_PORT '{port}' must be a valid integer") from err
 
@@ -1608,6 +1609,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_KV_CACHE_LAYOUT": env_with_choices(
         "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
     ),
+    # Opt into packed per-block KV cache allocation for multi-group
+    # attention-only HMA models (e.g. gpt-oss, Gemma 3/4).
+    "VLLM_USE_PACKED_HMA_KV_CACHE": lambda: bool(
+        int(os.getenv("VLLM_USE_PACKED_HMA_KV_CACHE", "0"))
+    ),
     # SSM conv state layout used for Mamba models.
     # - SD: (state_len, dim) — dim contiguous (default)
     # - DS: (dim, state_len) — TP-sharded dim on dim1,

vllm/lora/lora_model.py

@@ -17,7 +17,7 @@ from vllm.lora.utils import (
 )
 from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
 from vllm.model_executor.models.utils import WeightsMapper
-from vllm.utils.platform_utils import is_pin_memory_available
+from vllm.utils.torch_utils import PIN_MEMORY
 
 logger = init_logger(__name__)
 
@@ -126,7 +126,7 @@ class LoRAModel:
         skip_prefixes: list[str] | None = None,
     ) -> "LoRAModel":
         """Create a LoRAModel from a dictionary of tensors."""
-        pin_memory = str(device) == "cpu" and is_pin_memory_available()
+        pin_memory = str(device) == "cpu" and PIN_MEMORY
         loras: dict[str, LoRALayerWeights] = {}
         for tensor_name, tensor in tensors.items():
             if is_base_embedding_weights(tensor_name):

vllm/lora/lora_weights.py

@@ -7,7 +7,7 @@ import torch
 import torch.types
 
 from vllm.lora.peft_helper import PEFTHelper
-from vllm.utils.platform_utils import is_pin_memory_available
+from vllm.utils.torch_utils import PIN_MEMORY
 
 
 class LoRALayerWeights:
@@ -79,7 +79,7 @@ class LoRALayerWeights:
         dtype: torch.dtype,
         device: torch.types.Device,
     ) -> "LoRALayerWeights":
-        pin_memory = str(device) == "cpu" and is_pin_memory_available()
+        pin_memory = str(device) == "cpu" and PIN_MEMORY
         lora_a = torch.zeros(
             [rank, input_dim], dtype=dtype, device=device, pin_memory=pin_memory
         )

vllm/lora/model_manager.py

@@ -42,7 +42,7 @@ from vllm.model_executor.models.utils import PPMissingLayer
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.encoder_budget import MultiModalBudget
 from vllm.utils.cache import LRUCache
-from vllm.utils.platform_utils import is_pin_memory_available
+from vllm.utils.torch_utils import PIN_MEMORY
 
 logger = init_logger(__name__)
 
@@ -801,7 +801,7 @@ class LoRAModelManager:
         # 2. The weight packing above (e.g., pack_moe) may invalidate the
         # pin_memory allocation, so we execute it after packing.
 
-        pin_memory = str(lora_device) == "cpu" and is_pin_memory_available()
+        pin_memory = str(lora_device) == "cpu" and PIN_MEMORY
         if pin_memory:
             for lora in lora_model.loras.values():
                 if isinstance(lora.lora_a, list):

vllm/model_executor/layers/attention/mla_attention.py

@@ -1684,12 +1684,13 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
                 #  [[0, 0, 0, 0], [256, 256, 256, 256], [512, 512, 512, 512]]
                 # Note(simon): this is done in CPU because of downstream's
                 # of `to_list`.
-                chunk_starts = (
+                chunk_starts = torch.empty(
+                    num_chunks, num_prefills, dtype=torch.int32, pin_memory=True
+                ).copy_(
                     torch.arange(num_chunks, dtype=torch.int32)
+                    .multiply_(max_context_chunk)
                     .unsqueeze(1)
-                    .expand(-1, num_prefills)
-                    * max_context_chunk
-                ).pin_memory()
+                )
                 chunk_ends = torch.min(
                     context_lens_cpu.unsqueeze(0), chunk_starts + max_context_chunk
                 )
@@ -1746,12 +1747,13 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
                         )
                         * self.dcp_local_block_size
                     )
-                    local_chunk_starts = (
+                    local_chunk_starts = torch.empty(
+                        num_chunks, num_prefills, dtype=torch.int32, pin_memory=True
+                    ).copy_(
                         torch.arange(num_chunks, dtype=torch.int32)
+                        .multiply_(padded_local_max_context_chunk_across_ranks)
                         .unsqueeze(1)
-                        .expand(-1, num_prefills)
-                        * padded_local_max_context_chunk_across_ranks
-                    ).pin_memory()
+                    )
                     local_chunk_ends = torch.min(
                         padded_local_context_lens_cpu.unsqueeze(0),
                         local_chunk_starts

vllm/model_executor/layers/attention/mm_encoder_attention.py

@@ -28,6 +28,7 @@ from vllm.utils.flashinfer import (
     is_flashinfer_cudnn_fp8_prefill_attn_supported,
 )
 from vllm.utils.math_utils import round_up
+from vllm.utils.torch_utils import async_tensor_h2d
 from vllm.v1.attention.backends.fa_utils import get_flash_attn_version
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
 from vllm.v1.attention.ops.vit_attn_wrappers import (
@@ -311,7 +312,7 @@ class MMEncoderAttention(CustomOp):
                 )
                 cu_seqlens = np.concatenate([cu_seqlens_qko, cu_seqlens_v])
 
-        cu_seqlens = torch.from_numpy(cu_seqlens).to(device, non_blocking=True)
+        cu_seqlens = async_tensor_h2d(cu_seqlens, device=device)
         return cu_seqlens
 
     def __init__(

vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py

@@ -21,7 +21,6 @@ from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEQuantConfig,
 )
 from vllm.model_executor.layers.fused_moe.experts.triton_moe import TritonExperts
-from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
 from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import (
     dequantize_to_dtype,
 )
@@ -135,14 +134,6 @@ class Nvfp4QuantizationEmulationTritonExperts(TritonExperts):
             swizzle=False,
         )
 
-        hidden_states, _ = moe_kernel_quantize_input(
-            A=hidden_states,
-            A_scale=self.quant_config.a1_gscale,
-            quant_dtype="nvfp4",
-            per_act_token_quant=False,
-            quantization_emulation=True,
-        )
-
         # Activation quantization/dequantization is deferred to
         # `moe_kernel_quantize_input` in TritonExperts.apply.
         super().apply(

vllm/model_executor/layers/fused_moe/experts/ocp_mx_emulation_moe.py

@@ -21,7 +21,6 @@ from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEQuantConfig,
 )
 from vllm.model_executor.layers.fused_moe.experts.triton_moe import TritonExperts
-from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
 from vllm.model_executor.layers.quantization.utils.mxfp4_utils import dequant_mxfp4
 from vllm.model_executor.layers.quantization.utils.mxfp6_utils import dequant_mxfp6
 from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import (
@@ -155,16 +154,6 @@ class OCP_MXQuantizationEmulationTritonExperts(TritonExperts):
             w2, self.w2_scale_val, hidden_states.dtype
         )
 
-        # Apply activation QDQ if needed by the OCP MX scheme
-        hidden_states, _ = moe_kernel_quantize_input(
-            A=hidden_states,
-            A_scale=None,
-            quant_dtype=self.quant_config.quant_dtype,
-            per_act_token_quant=False,
-            ocp_mx_scheme=self.ocp_mx_scheme,
-            quantization_emulation=True,
-        )
-
         # Activation quantization/dequantization is deferred to
         # `moe_kernel_quantize_input` in TritonExperts.apply.
         super().apply(

vllm/model_executor/layers/fused_moe/experts/triton_moe.py

@@ -245,7 +245,7 @@ class TritonExperts(LoRAExpertsMixin, mk.FusedMoEExpertsModular):
             lora_unquantized_hidden_states = hidden_states
             hidden_states, a1q_scale = moe_kernel_quantize_input(
                 hidden_states,
-                self.a1_scale,
+                self.a1_scale or self.a1_gscale,
                 self.quant_dtype,
                 self.per_act_token_quant,
                 self.block_shape,

vllm/model_executor/layers/fused_moe/utils.py

@@ -296,6 +296,7 @@ def moe_kernel_quantize_input(
         if not quantization_emulation:
             return _nvfp4_quantize(A, A_scale, is_sf_swizzled_layout=is_scale_swizzled)
         else:
+            assert A_scale is not None
             A = ref_nvfp4_quant_dequant(A, A_scale, block_size=16)
             return A, None
     elif quant_dtype == "mxfp4":

vllm/model_executor/layers/pooler/seqwise/methods.py

@@ -10,6 +10,7 @@ import torch.nn as nn
 from vllm.config.pooler import SequencePoolingType
 from vllm.model_executor.layers.pooler import PoolingParamsUpdate
 from vllm.tasks import PoolingTask
+from vllm.utils.torch_utils import async_tensor_h2d
 from vllm.v1.pool.metadata import PoolingMetadata
 
 SequencePoolingMethodOutput: TypeAlias = torch.Tensor | list[torch.Tensor]
@@ -74,15 +75,14 @@ class MeanPool(SequencePoolingMethod):
             # early return for empty batch
             return hidden_states.new_empty((0, hidden_size), dtype=torch.float32)
 
-        # Build segment_ids on CPU so repeat_interleave doesn't need to sync
-        # GPU->CPU to learn its data-dependent output length, then upload
-        # non-blocking. eg. [2, 1, 3] -> [0, 0, 1, 2, 2, 2]
+        prompt_lens = async_tensor_h2d(
+            prompt_lens_cpu, device=hidden_states.device, dtype=torch.int64
+        )
+        # eg. [2, 1, 3] -> [0, 0, 1, 2, 2, 2]
         segment_ids = torch.repeat_interleave(
-            torch.arange(num_seqs, dtype=torch.long),
-            prompt_lens_cpu,
-        ).to(hidden_states.device, non_blocking=True)
-        prompt_lens = prompt_lens_cpu.to(
-            hidden_states.device, dtype=torch.int64, non_blocking=True
+            torch.arange(num_seqs, device=hidden_states.device, dtype=torch.long),
+            prompt_lens,
+            output_size=int(prompt_lens_cpu.sum()),
         )
         segment_sums = torch.zeros(
             (num_seqs, hidden_size),

vllm/model_executor/models/deepseek_v2.py

@@ -1001,24 +1001,30 @@ class DeepseekV2MLAAttention(nn.Module):
         # IndexCache config
         # Refer: https://arxiv.org/abs/2603.12201 for more details.
         _skip_topk = False
-        _index_topk_freq = getattr(config, "index_topk_freq", 1)
-        _index_topk_pattern = getattr(config, "index_topk_pattern", None)
-        _index_skip_topk_offset = getattr(config, "index_skip_topk_offset", 2)
-        layer_id = extract_layer_index(prefix)
+        is_mtp_layer = False
+        if self.is_v32:
+            _index_topk_freq = getattr(config, "index_topk_freq", 1)
+            _index_topk_pattern = getattr(config, "index_topk_pattern", None)
+            _index_skip_topk_offset = getattr(config, "index_skip_topk_offset", 2)
+            layer_id = extract_layer_index(prefix)
 
-        if _index_topk_pattern is None:
-            _skip_topk = (
-                max(layer_id - _index_skip_topk_offset + 1, 0) % _index_topk_freq != 0
+            if _index_topk_pattern is None:
+                _skip_topk = (
+                    max(layer_id - _index_skip_topk_offset + 1, 0) % _index_topk_freq
+                    != 0
+                )
+            elif 0 <= layer_id < len(_index_topk_pattern):
+                _skip_topk = _index_topk_pattern[layer_id] == "S"
+
+            # The skip pattern only governs backbone layers. MTP/nextn
+            # layers (layer_id >= num_hidden_layers) always build a full
+            # indexer: they compute indices at draft step 0 and toggle
+            # at runtime via set_skip_topk
+            # (index_share_for_mtp_iteration).
+            _num_hidden_layers = getattr(config, "num_hidden_layers", None)
+            is_mtp_layer = (
+                _num_hidden_layers is not None and layer_id >= _num_hidden_layers
             )
-        elif 0 <= layer_id < len(_index_topk_pattern):
-            _skip_topk = _index_topk_pattern[layer_id] == "S"
-
-        # The skip pattern only governs backbone layers. MTP/nextn layers
-        # (layer_id >= num_hidden_layers) always build a full indexer: they
-        # compute indices at draft step 0 and toggle at runtime via
-        # set_skip_topk (index_share_for_mtp_iteration).
-        _num_hidden_layers = getattr(config, "num_hidden_layers", None)
-        is_mtp_layer = _num_hidden_layers is not None and layer_id >= _num_hidden_layers
 
         if self.is_v32 and (not _skip_topk or is_mtp_layer):
             self.indexer_rope_emb = get_rope(

vllm/model_executor/models/moonvit.py

@@ -66,6 +66,7 @@ from vllm.model_executor.models.utils import maybe_prefix
 from vllm.model_executor.models.vision import is_vit_use_data_parallel
 from vllm.platforms import current_platform
 from vllm.transformers_utils.configs.moonvit import MoonViTConfig
+from vllm.utils.torch_utils import async_tensor_h2d
 
 
 def _apply_rope_input_validation(x, freqs_cis):
@@ -758,7 +759,7 @@ class MoonVitPretrainedModel(PreTrainedModel):
                         ),
                     ]
                 )
-        metadata["cu_seqlens"] = torch.from_numpy(cu_seqlens_np).to(device)
+        metadata["cu_seqlens"] = async_tensor_h2d(cu_seqlens_np, device=device)
 
         if max_seqlen_override is not None:
             max_seqlen_val = int(max_seqlen_override)
@@ -770,7 +771,7 @@ class MoonVitPretrainedModel(PreTrainedModel):
         metadata["max_seqlen"] = torch.tensor(max_seqlen_val, dtype=torch.int32)
 
         gather_idx_np = _build_merge_gather_idx(grid_pairs, self.merge_kernel_size)
-        metadata["merge_gather_idx"] = torch.from_numpy(gather_idx_np).to(device)
+        metadata["merge_gather_idx"] = async_tensor_h2d(gather_idx_np, device=device)
 
         return metadata
 

vllm/model_executor/models/qwen2_5_vl.py

@@ -83,9 +83,8 @@ from vllm.multimodal.parse import MultiModalDataItems
 from vllm.multimodal.processing import PromptReplacement, PromptUpdate
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
-from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
-from vllm.utils.torch_utils import async_tensor_h2d
+from vllm.utils.torch_utils import PIN_MEMORY, async_tensor_h2d
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
 from vllm.v1.worker.encoder_cudagraph_defs import EncoderCudaGraphReplayBuffers
 
@@ -825,7 +824,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
     @staticmethod
     def invert_permutation(perm: torch.Tensor) -> torch.Tensor:
         # building the inverse permutation in O(n) time
-        inv = torch.empty_like(perm, pin_memory=is_pin_memory_available())
+        inv = torch.empty_like(perm, pin_memory=PIN_MEMORY)
         inv[perm] = torch.arange(perm.numel(), device=perm.device, dtype=perm.dtype)
         return inv
 

vllm/model_executor/models/qwen3_vl.py

@@ -1202,6 +1202,49 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
         return video_items
 
 
+def _replace_video_token_placeholders(
+    prompt_ids: list[int],
+    target: list[int],
+    replacements: list[list[int]],
+) -> list[int]:
+    """Replace each 3-token video placeholder with its expanded sequence.
+
+    Args:
+        prompt_ids: Token IDs of the original (unexpanded) prompt.
+        target: 3-element list ``[vision_start_id, video_pad_id,
+            vision_end_id]`` to search for.
+        replacements: Per-video expanded token sequences, in prompt order.
+
+    Returns:
+        Token IDs with every placeholder triplet replaced.
+    """
+    result: list[int] = []
+    repl_idx = 0
+    i = 0
+    n = len(prompt_ids)
+    t0, t1, t2 = target
+    num_repl = len(replacements)
+
+    while i < n:
+        if (
+            i + 2 < n
+            and prompt_ids[i] == t0
+            and prompt_ids[i + 1] == t1
+            and prompt_ids[i + 2] == t2
+        ):
+            result.extend(replacements[repl_idx])
+            repl_idx += 1
+            i += 3
+        else:
+            result.append(prompt_ids[i])
+            i += 1
+
+    assert repl_idx == num_repl, (
+        f"Found {repl_idx} video placeholders but expected {num_repl}"
+    )
+    return result
+
+
 class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo]):
     def _call_hf_processor(
         self,
@@ -1211,15 +1254,23 @@ class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo])
         tok_kwargs: Mapping[str, object],
     ) -> BatchFeature:
         mm_data = dict(mm_data)
-        processor = self.info.get_hf_processor(**mm_kwargs)
 
         # Separate video processing from image processing. Because the videos
         # are processed into several image patches
+        video_input_ids_lst: list[list[int]] = []
         if videos := mm_data.pop("videos", []):
             video_grid_thw_lst = []
             pixel_values_videos_lst = []
             timestamps_per_video = []
 
+            hf_config = self.info.get_hf_config()
+            tokenizer = self.info.get_tokenizer()
+            merge_size = hf_config.vision_config.spatial_merge_size
+            video_pruning_rate = self.info.ctx.get_mm_config().video_pruning_rate
+            vision_start_token_id = hf_config.vision_start_token_id
+            vision_end_token_id = hf_config.vision_end_token_id
+            video_token_id = hf_config.video_token_id
+
             for item in videos:
                 video_array, metadata = item
 
@@ -1269,55 +1320,38 @@ class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo])
                     tok_kwargs=tok_kwargs,
                 )
 
-                merge_size = processor.video_processor.merge_size
-                # Get video grid info for EVS calculation.
+                # Discard HF output input_ids — we use get_video_repl below
+                # to generate the correct (EVS-adjusted) token sequence.
+                video_outputs.pop("input_ids", None)
+
                 video_grid_thw = video_outputs["video_grid_thw"]
                 num_frames = int(video_grid_thw[0, 0])
                 tokens_per_frame_base = int(video_grid_thw[0, 1:].prod()) // (
                     merge_size**2
                 )
 
-                # Apply EVS if enabled.
-                video_pruning_rate = self.info.ctx.get_mm_config().video_pruning_rate
                 if video_pruning_rate is not None and video_pruning_rate > 0.0:
                     num_tokens = compute_retained_tokens_count(
                         tokens_per_frame=tokens_per_frame_base,
                         num_frames=num_frames,
                         q=video_pruning_rate,
                     )
-                    # Here we just need placeholders that won't actually be replaced -
-                    # we just need to make sure the total number of tokens is correct
-                    # assign all tokens to the first frame.
                     tokens_per_frame = [num_tokens] + [0] * (num_frames - 1)
                     select_token_id = False
                 else:
                     tokens_per_frame = [tokens_per_frame_base] * num_frames
                     select_token_id = True
 
-                # Generate the video replacement with EVS-adjusted token counts
-                tokenizer = self.info.get_tokenizer()
-                hf_config = self.info.get_hf_config()
                 video_repl = Qwen3VLMultiModalProcessor.get_video_repl(
                     tokens_per_frame=tokens_per_frame,
                     timestamps=timestamps,
                     tokenizer=tokenizer,
-                    vision_start_token_id=hf_config.vision_start_token_id,
-                    vision_end_token_id=hf_config.vision_end_token_id,
-                    video_token_id=hf_config.video_token_id,
+                    vision_start_token_id=vision_start_token_id,
+                    vision_end_token_id=vision_end_token_id,
+                    video_token_id=video_token_id,
                     select_token_id=select_token_id,
                 )
-
-                # Convert token IDs to text for the HF processor flow
-                video_placeholder = tokenizer.decode(
-                    video_repl.full, skip_special_tokens=False
-                )
-                input_ids = video_outputs.pop("input_ids")
-                video_placeholder = processor.tokenizer.batch_decode(input_ids)[0]
-                prompt = prompt.replace(
-                    "<|vision_start|><|video_pad|><|vision_end|>",
-                    video_placeholder,
-                    1,
-                )
+                video_input_ids_lst.append(list(video_repl.full))
 
                 video_grid_thw_lst.append(video_outputs["video_grid_thw"])
                 pixel_values_videos_lst.append(video_outputs["pixel_values_videos"])
@@ -1335,6 +1369,24 @@ class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo])
             mm_kwargs=mm_kwargs,
             tok_kwargs=tok_kwargs,
         )
+
+        # Replace each placeholder triplet with pre-computed video tokens.
+        if video_input_ids_lst:
+            hf_config = self.info.get_hf_config()
+            video_target = [
+                hf_config.vision_start_token_id,
+                hf_config.video_token_id,
+                hf_config.vision_end_token_id,
+            ]
+            input_ids = processed_outputs.pop("input_ids")
+            if not isinstance(input_ids, list):
+                input_ids = input_ids.tolist()
+            (prompt_ids,) = input_ids
+            expanded_ids = _replace_video_token_placeholders(
+                prompt_ids, video_target, video_input_ids_lst
+            )
+            processed_outputs["input_ids"] = [expanded_ids]
+
         combined_outputs = dict(
             processed_outputs,
             **video_outputs,

vllm/models/deepseek_v4/sparse_mla.py

@@ -13,6 +13,7 @@ from vllm.config.cache import CacheDType
 from vllm.platforms.interface import DeviceCapability
 from vllm.triton_utils import tl, triton
 from vllm.utils.math_utils import cdiv
+from vllm.utils.torch_utils import np_to_pinned_tensor
 from vllm.v1.attention.backend import (
     AttentionBackend,
     AttentionCGSupport,
@@ -207,7 +208,7 @@ class DeepseekV4FlashMLAMetadataBuilder(
         # Zero-fill for cudagraphs
         self.req_id_per_token_buffer.fill_(0)
         self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_(
-            torch.from_numpy(req_id_per_token), non_blocking=True
+            np_to_pinned_tensor(req_id_per_token), non_blocking=True
         )
         req_id_per_token = self.req_id_per_token_buffer[:num_tokens]
 

vllm/multimodal/inputs.py

@@ -488,7 +488,13 @@ class MultiModalBatchedField(BaseMultiModalField):
                 # An optimization when `batch` contains only one tensor:
                 # - produce exactly same result as `torch.stack(batch)`
                 # - will achieve zero-copy if the tensor is contiguous
-                return batch[0].unsqueeze(0).contiguous()
+                out = batch[0].unsqueeze(0)
+                if not pin_memory:
+                    return out.contiguous()
+                # Avoid extra copy - pinning unpinned memory will make it contiguous
+                if not out.is_contiguous() and out.is_pinned():
+                    out = out.contiguous()
+                return out.pin_memory()
             first_shape = batch[0].shape
             if all(elem.shape == first_shape for elem in batch):
                 out = torch.empty(
@@ -538,7 +544,13 @@ class MultiModalFlatField(BaseMultiModalField):
                 # An optimization when `batch` contains only one tensor:
                 # - produce exactly same result as `torch.concat(batch)`
                 # - will achieve zero-copy if the tensor is contiguous
-                return batch[0].contiguous()
+                out = batch[0]
+                if not pin_memory:
+                    return out.contiguous()
+                # Avoid extra copy - pinning unpinned memory will make it contiguous
+                if not out.is_contiguous() and out.is_pinned():
+                    out = out.contiguous()
+                return out.pin_memory()
 
             dim = self.dim + (self.dim < 0) * len(batch[0].shape)
 

vllm/parser/minimax_m3.py

@@ -0,0 +1,163 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""MiniMax M3 parser for reasoning markers."""
+
+from __future__ import annotations
+
+import functools
+from collections.abc import Iterable, Sequence
+from typing import TYPE_CHECKING
+
+from vllm.parser.engine.events import EventType
+from vllm.parser.engine.parser_engine import ParserEngine
+from vllm.parser.engine.parser_engine_config import (
+    ParserEngineConfig,
+    ParserState,
+    Transition,
+)
+
+if TYPE_CHECKING:
+    from vllm.tokenizers import TokenizerLike
+    from vllm.tool_parsers.abstract_tool_parser import Tool
+
+THINK_START = "<mm:think>"
+THINK_END = "</mm:think>"
+
+
+@functools.cache
+def minimax_m3_config(thinking: bool = False) -> ParserEngineConfig:
+    return ParserEngineConfig(
+        name="minimax_m3",
+        initial_state=ParserState.REASONING if thinking else ParserState.CONTENT,
+        terminals={
+            "THINK_START": THINK_START,
+            "THINK_END": THINK_END,
+        },
+        transitions={
+            (ParserState.CONTENT, "THINK_START"): Transition(
+                ParserState.REASONING,
+                (EventType.REASONING_START,),
+            ),
+            (ParserState.REASONING, "THINK_START"): Transition(
+                ParserState.REASONING,
+                (),
+            ),
+            (ParserState.REASONING, "THINK_END"): Transition(
+                ParserState.CONTENT,
+                (EventType.REASONING_END,),
+            ),
+            (ParserState.CONTENT, "THINK_END"): Transition(
+                ParserState.CONTENT,
+                (),
+            ),
+        },
+    )
+
+
+class MiniMaxM3Parser(ParserEngine):
+    """MiniMax M3 parser backed by the declarative parser engine."""
+
+    def __init__(
+        self,
+        tokenizer: TokenizerLike,
+        tools: list[Tool] | None = None,
+        **kwargs,
+    ) -> None:
+        chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
+        self._initial_in_reasoning = chat_kwargs.get("thinking_mode") == "enabled"
+        kwargs.setdefault(
+            "parser_engine_config",
+            minimax_m3_config(thinking=self._initial_in_reasoning),
+        )
+        super().__init__(tokenizer, tools, **kwargs)
+        self._start_token_ids = self._encode_marker(THINK_START)
+        self._end_token_ids = self._encode_marker(THINK_END)
+
+    def _encode_marker(self, marker: str) -> tuple[int, ...]:
+        try:
+            token_ids = self.model_tokenizer.encode(marker, add_special_tokens=False)
+        except TypeError:
+            token_ids = self.model_tokenizer.encode(marker)
+        return tuple(token_ids)
+
+    @staticmethod
+    def _contains_token_sequence(
+        token_ids: Sequence[int], marker_ids: Sequence[int]
+    ) -> bool:
+        if not marker_ids or len(marker_ids) > len(token_ids):
+            return False
+        marker_len = len(marker_ids)
+        return any(
+            tuple(token_ids[i : i + marker_len]) == tuple(marker_ids)
+            for i in range(len(token_ids) - marker_len + 1)
+        )
+
+    @staticmethod
+    def _rfind_token_sequence(
+        token_ids: Sequence[int], marker_ids: Sequence[int]
+    ) -> int:
+        if not marker_ids or len(marker_ids) > len(token_ids):
+            return -1
+        marker_len = len(marker_ids)
+        for i in range(len(token_ids) - marker_len, -1, -1):
+            if tuple(token_ids[i : i + marker_len]) == tuple(marker_ids):
+                return i
+        return -1
+
+    def is_reasoning_end(self, input_ids: list[int]) -> bool:
+        start_index = self._rfind_token_sequence(input_ids, self._start_token_ids)
+        end_index = self._rfind_token_sequence(input_ids, self._end_token_ids)
+        if end_index < 0:
+            return False
+        if start_index < 0:
+            return True
+        return end_index > start_index
+
+    def is_reasoning_end_streaming(
+        self, input_ids: Sequence[int], delta_ids: Iterable[int]
+    ) -> bool:
+        if self.reasoning_ended:
+            return True
+        if self._engine._lexer.buffer:
+            return False
+        if self._initial_in_reasoning:
+            return False
+        if self._engine.state == ParserState.CONTENT:
+            return bool(input_ids)
+        return False
+
+    def extract_content_ids(self, input_ids: list[int]) -> list[int]:
+        end_index = self._rfind_token_sequence(input_ids, self._end_token_ids)
+        if end_index >= 0:
+            return input_ids[end_index + len(self._end_token_ids) :]
+
+        has_start = self._contains_token_sequence(input_ids, self._start_token_ids)
+        if self._initial_in_reasoning and not has_start:
+            return []
+
+        if not has_start:
+            return input_ids
+        return []
+
+    def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int:
+        count = 0
+        depth = 1 if self._initial_in_reasoning else 0
+        i = 0
+        while i < len(token_ids):
+            if tuple(token_ids[i : i + len(self._start_token_ids)]) == (
+                self._start_token_ids
+            ):
+                depth += 1
+                i += len(self._start_token_ids)
+                continue
+            if tuple(token_ids[i : i + len(self._end_token_ids)]) == (
+                self._end_token_ids
+            ):
+                if depth > 0:
+                    depth -= 1
+                i += len(self._end_token_ids)
+                continue
+            if depth > 0:
+                count += 1
+            i += 1
+        return count

vllm/platforms/__init__.py

@@ -9,7 +9,6 @@ from typing import TYPE_CHECKING
 from vllm import envs
 from vllm.plugins import PLATFORM_PLUGINS_GROUP, load_plugins_by_group
 from vllm.utils.import_utils import resolve_obj_by_qualname
-from vllm.utils.torch_utils import supports_xccl
 
 from .interface import CpuArchEnum, Platform, PlatformEnum
 
@@ -135,7 +134,7 @@ def xpu_platform_plugin() -> str | None:
     try:
         import torch
 
-        if supports_xccl():
+        if torch.distributed.is_xccl_available():
             dist_backend = "xccl"
             from vllm.platforms.xpu import XPUPlatform
 

vllm/platforms/cuda.py

@@ -23,7 +23,6 @@ import vllm._C_stable_libtorch  # noqa
 import vllm.envs as envs
 from vllm.logger import init_logger
 from vllm.utils.import_utils import import_pynvml
-from vllm.utils.torch_utils import is_quantized_kv_cache
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
 from .interface import DeviceCapability, Platform, PlatformEnum, in_wsl
@@ -88,6 +87,8 @@ def _get_backend_priorities(
     kv_cache_dtype: CacheDType | None = None,
 ) -> list[AttentionBackendEnum]:
     """Get backend priorities with lazy import to avoid circular dependency."""
+    from vllm.utils.torch_utils import is_quantized_kv_cache
+
     if use_mla:
         if device_capability.major == 10:
             # Sparse MLA backend priorities
@@ -685,6 +686,15 @@ class CudaPlatformBase(Platform):
 # all the related functions work on real physical device ids.
 # the major benefit of using NVML is that it will not initialize CUDA
 class NvmlCudaPlatform(CudaPlatformBase):
+    @classmethod
+    @with_nvml_context
+    def device_control_id_to_physical_device_id(cls, device_id: str) -> int:
+        try:
+            return int(device_id)
+        except ValueError:
+            handle = pynvml.nvmlDeviceGetHandleByUUID(device_id)
+            return pynvml.nvmlDeviceGetIndex(handle)
+
     @classmethod
     @cache
     @with_nvml_context

vllm/platforms/interface.py

@@ -30,6 +30,33 @@ else:
 
 logger = init_logger(__name__)
 
+_assigned_physical_gpu_ids: list[int] | None = None
+
+
+def set_assigned_physical_gpu_ids(ids: list[int]) -> None:
+    """Set the physical GPU IDs assigned to this worker process.
+    Called during worker init so that device_id_to_physical_device_id()
+    can map local_rank to the correct physical device without relying
+    on CUDA_VISIBLE_DEVICES.
+
+    Idempotent: a second call with the same value is a no-op.
+    Raises RuntimeError if called again with a different value.
+
+    This is expected to run during single-threaded worker initialization."""
+    global _assigned_physical_gpu_ids
+    if _assigned_physical_gpu_ids is not None:
+        if _assigned_physical_gpu_ids != ids:
+            raise RuntimeError(
+                f"set_assigned_physical_gpu_ids called with conflicting values: "
+                f"existing={_assigned_physical_gpu_ids}, new={ids}"
+            )
+        return
+    _assigned_physical_gpu_ids = ids
+
+
+def get_assigned_physical_gpu_ids() -> list[int] | None:
+    return _assigned_physical_gpu_ids
+
 
 @functools.cache
 def in_wsl() -> bool:
@@ -233,8 +260,34 @@ class Platform:
         """
         import vllm.kernels  # noqa: F401
 
+    @classmethod
+    def device_control_id_to_physical_device_id(cls, device_id: str) -> int:
+        """Map one device-control env entry to an integer physical device ID."""
+        try:
+            return int(device_id)
+        except ValueError as e:
+            raise ValueError(
+                f"Non-integer device ID {device_id!r} is not supported by "
+                f"{cls.device_name}."
+            ) from e
+
     @classmethod
     def device_id_to_physical_device_id(cls, device_id: int):
+        """Map a vLLM-local logical device ID to a physical device ID.
+
+        The input is a logical local ID (e.g. a local rank), NOT a visible
+        device ordinal; for the latter use
+        visible_device_id_to_physical_device_id(). The two coincide only
+        when no logical-to-physical mapping is in effect.
+        """
+        if _assigned_physical_gpu_ids is not None:
+            if device_id >= len(_assigned_physical_gpu_ids):
+                raise IndexError(
+                    f"device_id {device_id} is out of range for "
+                    f"assigned_physical_gpu_ids {_assigned_physical_gpu_ids} "
+                    f"({len(_assigned_physical_gpu_ids)} devices assigned)"
+                )
+            return _assigned_physical_gpu_ids[device_id]
         # Treat empty device control env var as unset. This is a valid
         # configuration in Ray setups where the engine is launched in
         # a CPU-only placement group located on a GPU node.
@@ -244,10 +297,58 @@ class Platform:
         ):
             device_ids = os.environ[cls.device_control_env_var].split(",")
             physical_device_id = device_ids[device_id]
-            return int(physical_device_id)
+            return cls.device_control_id_to_physical_device_id(physical_device_id)
         else:
             return device_id
 
+    @classmethod
+    def logical_device_id_to_visible_device_id(cls, device_id: int) -> int:
+        """Map a vLLM-local logical device ID to the current process's
+        visible accelerator ordinal.
+
+        vLLM internals use logical local IDs. Physical IDs are used only
+        at platform/topology boundaries. This helper performs the final
+        translation needed by APIs such as ``torch.device("cuda:N")``.
+        """
+        physical_device_id = cls.device_id_to_physical_device_id(device_id)
+        device_control_env = os.environ.get(cls.device_control_env_var, "")
+        if not device_control_env:
+            return physical_device_id
+
+        visible_physical_device_ids = [
+            cls.device_control_id_to_physical_device_id(physical_id)
+            for physical_id in device_control_env.split(",")
+        ]
+        if physical_device_id not in visible_physical_device_ids:
+            raise RuntimeError(
+                f"Physical device {physical_device_id} for logical device "
+                f"{device_id} is not visible in {cls.device_control_env_var}="
+                f"{device_control_env}"
+            )
+        return visible_physical_device_ids.index(physical_device_id)
+
+    @classmethod
+    def visible_device_id_to_physical_device_id(cls, device_id: int) -> int:
+        """Map a visible accelerator ordinal (e.g. ``torch.device.index``)
+        to a physical device ID.
+
+        This is the inverse of the env-var translation performed by
+        logical_device_id_to_visible_device_id() and is independent of any
+        logical-to-physical mapping set via set_assigned_physical_gpu_ids().
+        """
+        device_control_env = os.environ.get(cls.device_control_env_var, "")
+        if not device_control_env:
+            return device_id
+        visible_device_ids = device_control_env.split(",")
+        if device_id >= len(visible_device_ids):
+            raise IndexError(
+                f"visible device ordinal {device_id} is out of range for "
+                f"{cls.device_control_env_var}={device_control_env}"
+            )
+        return cls.device_control_id_to_physical_device_id(
+            visible_device_ids[device_id]
+        )
+
     @classmethod
     def import_kernels(cls) -> None:
         """Import any platform-specific C kernels."""

vllm/platforms/xpu.py

@@ -14,7 +14,6 @@ import vllm_xpu_kernels._xpu_C  # noqa
 
 import vllm.envs as envs
 from vllm.logger import init_logger
-from vllm.utils.torch_utils import supports_xpu_graph
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
 from .interface import DeviceCapability, Platform, PlatformEnum
@@ -178,8 +177,6 @@ class XPUPlatform(Platform):
 
     @classmethod
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
-        parallel_config = vllm_config.parallel_config
-
         # lazy import to avoid circular import
         from vllm.config import CUDAGraphMode
 
@@ -190,6 +187,10 @@ class XPUPlatform(Platform):
         attention_config = vllm_config.attention_config
         if attention_config.backend is None:
             attention_config.backend = AttentionBackendEnum.FLASH_ATTN
+
+        # lazy import to avoid circular import
+        from vllm.utils.torch_utils import supports_xpu_graph
+
         if not supports_xpu_graph():
             compilation_config.cudagraph_mode = CUDAGraphMode.NONE
             logger.warning(
@@ -324,9 +325,8 @@ class XPUPlatform(Platform):
 
     @classmethod
     def get_device_communicator_cls(cls) -> str:
-        from vllm.utils.torch_utils import supports_xccl
-
-        if not supports_xccl():
+        if not torch.distributed.is_xccl_available():
+            # Supports xccl with PyTorch versions >= 2.8.0.dev for XPU platform
             logger.warning(
                 "xccl is not enabled in this torch build, communication"
                 " is not available."

vllm/reasoning/minimax_m3_reasoning_parser.py

@@ -2,170 +2,19 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from collections.abc import Iterable, Sequence
-from typing import TYPE_CHECKING
 
-from vllm.entrypoints.openai.engine.protocol import DeltaMessage
-from vllm.reasoning.basic_parsers import BaseThinkingReasoningParser
-
-if TYPE_CHECKING:
-    from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
-    from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
+from vllm.parser.engine.adapters import ParserEngineReasoningAdapter
+from vllm.parser.minimax_m3 import MiniMaxM3Parser
 
 
-class MiniMaxM3ReasoningParser(BaseThinkingReasoningParser):
-    """Reasoning parser for MiniMax M3 explicit thinking blocks.
+class MiniMaxM3ReasoningParser(ParserEngineReasoningAdapter):
+    """Reasoning parser adapter for MiniMax M3 explicit thinking blocks."""
 
-    MiniMax M3 emits reasoning as:
-
-        <mm:think>reasoning text</mm:think>assistant content
-
-    The M3 tokenizer exposes both markers as complete vocabulary tokens. The
-    chat template may also prefill the start marker when
-    ``thinking_mode="enabled"``, so generated text can begin directly inside a
-    reasoning block without emitting ``<mm:think>`` again.
-    """
-
-    @property
-    def start_token(self) -> str:
-        return "<mm:think>"
-
-    @property
-    def end_token(self) -> str:
-        return "</mm:think>"
-
-    def __init__(self, tokenizer, *args, **kwargs):
-        super().__init__(tokenizer, *args, **kwargs)
-        chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
-        self._initial_in_reasoning = chat_kwargs.get("thinking_mode") == "enabled"
-        self._at_response_start = True
-
-    def extract_reasoning(
-        self,
-        model_output: str,
-        request: "ChatCompletionRequest | ResponsesRequest",
-    ) -> tuple[str | None, str | None]:
-        # MiniMax M3 can start a response with a stray closer. Drop that first
-        # token only; later unmatched closers stay visible as content.
-        if not self._initial_in_reasoning and model_output.startswith(self.end_token):
-            content = model_output[len(self.end_token) :]
-            return None, content or None
-
-        if self._initial_in_reasoning and self.start_token not in model_output:
-            reasoning, end, content = model_output.partition(self.end_token)
-            if not end:
-                return model_output, None
-            return reasoning, content or None
-
-        if self.start_token not in model_output:
-            return None, model_output
-
-        content_before, _, after_start = model_output.partition(self.start_token)
-        reasoning, end, content_after = after_start.partition(self.end_token)
-        if not end:
-            return reasoning, content_before or None
-
-        return reasoning, (content_before + content_after) or None
+    _parser_engine_cls = MiniMaxM3Parser
 
     def is_reasoning_end_streaming(
         self, input_ids: Sequence[int], delta_ids: Iterable[int]
     ) -> bool:
-        delta_ids = tuple(delta_ids)
-        if self.end_token_id in delta_ids:
-            return True
-        if self.end_token_id in input_ids:
-            return True
-        if self._initial_in_reasoning:
-            return False
-        if self.start_token_id not in input_ids:
-            return bool(input_ids)
-        return False
-
-    def extract_content_ids(self, input_ids: list[int]) -> list[int]:
-        if self.end_token_id in input_ids:
-            end_index = len(input_ids) - 1 - input_ids[::-1].index(self.end_token_id)
-            return input_ids[end_index + 1 :]
-
-        if self._initial_in_reasoning and self.start_token_id not in input_ids:
-            return []
-
-        if self.start_token_id not in input_ids:
-            return input_ids
-        return []
-
-    def extract_reasoning_streaming(
-        self,
-        previous_text: str,
-        current_text: str,
-        delta_text: str,
-        previous_token_ids: Sequence[int],
-        current_token_ids: Sequence[int],
-        delta_token_ids: Sequence[int],
-    ) -> DeltaMessage | None:
-        if not delta_text:
-            return None
-
-        if self._at_response_start and not self._initial_in_reasoning:
-            # Apply the leading-closer tolerance once. Later unmatched closers
-            # stay visible as content.
-            self._at_response_start = False
-            if delta_text.startswith(self.end_token):
-                delta_text = delta_text[len(self.end_token) :]
-                if not delta_text:
-                    return None
-                if delta_token_ids and delta_token_ids[0] == self.end_token_id:
-                    delta_token_ids = delta_token_ids[1:]
-
-        if self.end_token_id in previous_token_ids:
-            return DeltaMessage(content=delta_text)
-
-        if (
-            self._initial_in_reasoning
-            and self.start_token_id not in previous_token_ids
-            and self.start_token_id not in delta_token_ids
-        ):
-            if self.end_token_id in delta_token_ids:
-                reasoning, _, content = delta_text.partition(self.end_token)
-                return DeltaMessage(
-                    reasoning=reasoning or None,
-                    content=content or None,
-                )
-            return DeltaMessage(reasoning=delta_text)
-
-        if (
-            self.start_token_id not in previous_token_ids
-            and self.start_token_id not in delta_token_ids
-        ):
-            return DeltaMessage(content=delta_text)
-
-        if self.end_token_id in delta_token_ids:
-            reasoning_text, _, content = delta_text.partition(self.end_token)
-            if self.start_token_id in delta_token_ids:
-                _, _, reasoning_text = reasoning_text.partition(self.start_token)
-            return DeltaMessage(
-                reasoning=reasoning_text or None,
-                content=content or None,
-            )
-
-        if self.start_token_id in delta_token_ids:
-            _, _, reasoning = delta_text.partition(self.start_token)
-            return DeltaMessage(reasoning=reasoning) if reasoning else None
-
-        return DeltaMessage(reasoning=delta_text)
-
-    def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int:
-        if not self._initial_in_reasoning:
-            return super().count_reasoning_tokens(token_ids)
-
-        count = 0
-        depth = 1
-        for token_id in token_ids:
-            if token_id == self.start_token_id:
-                depth += 1
-                continue
-            if token_id == self.end_token_id:
-                if depth > 0:
-                    depth -= 1
-                continue
-            if depth > 0:
-                count += 1
-        return count
+        return self._parser_engine.is_reasoning_end_streaming(
+            list(input_ids), tuple(delta_ids)
+        )

vllm/tool_parsers/__init__.py

@@ -15,7 +15,7 @@ Register a lazy module mapping.
 Example:
     ToolParserManager.register_lazy_module(
         name="kimi_k2",
-        module_path="vllm.tool_parsers.kimi_k2_parser",
+        module_path="vllm.tool_parsers.kimi_k2_tool_parser",
         class_name="KimiK2ToolParser",
     )
 """

vllm/utils/torch_utils.py

@@ -3,7 +3,6 @@
 import contextlib
 import importlib.metadata
 import os
-import platform
 import random
 import threading
 from collections.abc import Callable, Collection
@@ -18,6 +17,7 @@ from torch.library import Library, infer_schema
 
 import vllm.envs as envs
 from vllm.logger import init_logger
+from vllm.utils.platform_utils import is_pin_memory_available
 
 if TYPE_CHECKING:
     from vllm.config import ModelConfig
@@ -68,9 +68,7 @@ MODELOPT_TO_VLLM_KV_CACHE_DTYPE_MAP = {
 T = TypeVar("T")
 
 
-# Pin memory in non-WSL case.
-# Logic duplicated here for now to avoid circular import.
-PIN_MEMORY = "microsoft" not in " ".join(platform.uname()).lower()
+PIN_MEMORY = is_pin_memory_available()
 
 
 def is_quantized_kv_cache(kv_cache_dtype: str) -> bool:
@@ -606,14 +604,24 @@ def create_kv_caches_with_random(
 
 
 def async_tensor_h2d(
-    data: list,
-    dtype: torch.dtype,
+    data: list | np.ndarray | torch.Tensor,
     device: str | torch.device,
-    pin_memory: bool = PIN_MEMORY,
+    dtype: torch.dtype | None = None,
 ) -> torch.Tensor:
-    """Asynchronously create a tensor and copy it from host to device."""
-    t = torch.tensor(data, dtype=dtype, pin_memory=pin_memory, device="cpu")
-    return t.to(device=device, non_blocking=True)
+    """Copy list/numpy array/tensor async from host to device."""
+    if isinstance(data, np.ndarray):
+        data = torch.from_numpy(data)
+    if isinstance(data, torch.Tensor):
+        t = data.pin_memory() if PIN_MEMORY else data
+    else:
+        t = torch.tensor(data, dtype=dtype, pin_memory=PIN_MEMORY, device="cpu")
+    assert t.is_cpu
+    return t.to(device=device, dtype=dtype, non_blocking=True)
+
+
+def np_to_pinned_tensor(array: np.ndarray) -> torch.Tensor:
+    t = torch.from_numpy(array)
+    return t.pin_memory() if PIN_MEMORY else t
 
 
 def make_ndarray_with_pad(
@@ -914,11 +922,6 @@ def _encode_layer_name(layer_name: str) -> str | LayerName:
     return LayerName(layer_name) if _USE_LAYERNAME else layer_name
 
 
-# Supports xccl with PyTorch versions >= 2.8.0.dev for XPU platform
-def supports_xccl() -> bool:
-    return torch.distributed.is_xccl_available()
-
-
 # Supports XPU Graph with PyTorch versions >= 2.11.0.dev for XPU platform
 def supports_xpu_graph() -> bool:
     return is_torch_equal_or_newer("2.11.0.dev")

vllm/v1/attention/backends/flashinfer.py

@@ -41,8 +41,8 @@ from vllm.utils.flashinfer import (
     use_trtllm_attention,
 )
 from vllm.utils.math_utils import cdiv
-from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.torch_utils import (
+    PIN_MEMORY,
     canonicalize_singleton_dim_strides,
     is_quantized_kv_cache,
     is_strictly_contiguous,
@@ -708,9 +708,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
         # Since we do not have explicit synchronization in ModelRunnerV2, we do not pin
         # reused CPU buffers to avoid a race condition between step N async copies to
         # GPU and step N+1 buffer updates.
-        self.pin_memory = (
-            not vllm_config.use_v2_model_runner and is_pin_memory_available()
-        )
+        self.pin_memory = not vllm_config.use_v2_model_runner and PIN_MEMORY
         self.paged_kv_indptr = self._make_buffer(max_num_reqs + 1)
         self.paged_kv_indptr_cpu_buffer = torch.zeros_like(
             self.paged_kv_indptr.cpu, pin_memory=self.pin_memory

vllm/v1/attention/backends/flex_attention.py

@@ -28,7 +28,11 @@ from vllm.logger import init_logger
 from vllm.model_executor.layers.attention import Attention
 from vllm.platforms import current_platform
 from vllm.utils.math_utils import cdiv
-from vllm.utils.torch_utils import is_quantized_kv_cache, is_torch_equal_or_newer
+from vllm.utils.torch_utils import (
+    async_tensor_h2d,
+    is_quantized_kv_cache,
+    is_torch_equal_or_newer,
+)
 from vllm.v1.attention.backend import (
     AttentionBackend,
     AttentionCGSupport,
@@ -58,7 +62,7 @@ def _offsets_to_doc_ids_tensor(
     doc_ids = torch.repeat_interleave(
         torch.arange(len(counts), dtype=torch.int32), counts
     )
-    return doc_ids.to(device, non_blocking=True)
+    return async_tensor_h2d(doc_ids, device=device)
 
 
 def pad_to_multiple(x: torch.Tensor, multiple: int, dim: int):

vllm/v1/attention/backends/gdn_attn.py

@@ -8,6 +8,7 @@ from typing import Literal
 import torch
 
 from vllm.config import VllmConfig
+from vllm.utils.torch_utils import async_tensor_h2d
 from vllm.v1.attention.backend import (
     AttentionBackend,
     AttentionCGSupport,
@@ -203,8 +204,8 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
                 spec_sequence_masks = None
                 spec_sequence_masks_cpu = None
             else:
-                spec_sequence_masks = spec_sequence_masks_cpu.to(
-                    query_start_loc.device, non_blocking=True
+                spec_sequence_masks = async_tensor_h2d(
+                    spec_sequence_masks_cpu, device=query_start_loc.device
                 )
 
         if spec_sequence_masks is None:
@@ -376,12 +377,14 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
                 )
 
                 assert prefill_query_start_loc_cpu is not None
-                chunk_indices = prepare_chunk_indices(
-                    prefill_query_start_loc_cpu, FLA_CHUNK_SIZE
-                ).to(device=gpu_device, non_blocking=True)
-                chunk_offsets = prepare_chunk_offsets(
-                    prefill_query_start_loc_cpu, FLA_CHUNK_SIZE
-                ).to(device=gpu_device, non_blocking=True)
+                chunk_indices = async_tensor_h2d(
+                    prepare_chunk_indices(prefill_query_start_loc_cpu, FLA_CHUNK_SIZE),
+                    device=gpu_device,
+                )
+                chunk_offsets = async_tensor_h2d(
+                    prepare_chunk_offsets(prefill_query_start_loc_cpu, FLA_CHUNK_SIZE),
+                    device=gpu_device,
+                )
 
         if num_prefills > 0:
             has_initial_state = context_lens_tensor > 0

vllm/v1/attention/backends/mamba2_attn.py

@@ -7,6 +7,7 @@ from typing import Any
 import torch
 
 from vllm.config import VllmConfig
+from vllm.utils.torch_utils import async_tensor_h2d
 from vllm.v1.attention.backend import (
     AttentionBackend,
     CommonAttentionMetadata,
@@ -68,22 +69,22 @@ def compute_varlen_chunk_metadata(
 
     # Exclusive prefix sum over logical-chunk lengths
     if chunk_lens:
-        cu_chunk_seqlens = torch.tensor(
-            [0] + list(itertools.accumulate(chunk_lens)),
-            device=device,
-            dtype=torch.int32,
-        )
-        # Final boundary must equal total tokens
-        assert int(cu_chunk_seqlens[-1].item()) == total
+        cu_chunk_seqlens_list = [0] + list(itertools.accumulate(chunk_lens))
+        # Final boundary must equal total tokens (check on host to avoid a sync)
+        assert cu_chunk_seqlens_list[-1] == total
     else:
-        cu_chunk_seqlens = torch.tensor([0], device=device, dtype=torch.int32)
+        cu_chunk_seqlens_list = [0]
+    cu_chunk_seqlens = async_tensor_h2d(
+        cu_chunk_seqlens_list, dtype=torch.int32, device=device
+    )
 
-    last_chunk_indices_t = (
-        torch.tensor(last_chunk_indices, device=device, dtype=torch.int32)
-        if len(starts) > 0
-        else torch.empty((0,), device=device, dtype=torch.int32)
+    # last_chunk_indices is empty when there are no sequences (len(starts) == 0).
+    last_chunk_indices_t = async_tensor_h2d(
+        last_chunk_indices, dtype=torch.int32, device=device
+    )
+    seq_idx_chunks_t = async_tensor_h2d(
+        seq_idx_chunks, dtype=torch.int32, device=device
     )
-    seq_idx_chunks_t = torch.tensor(seq_idx_chunks, device=device, dtype=torch.int32)
     return cu_chunk_seqlens, last_chunk_indices_t, seq_idx_chunks_t
 
 

vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py

@@ -26,7 +26,7 @@ from vllm.model_executor.layers.attention.mla_attention import (
     get_mla_dims,
 )
 from vllm.platforms.interface import DeviceCapability
-from vllm.utils.torch_utils import is_quantized_kv_cache
+from vllm.utils.torch_utils import is_quantized_kv_cache, np_to_pinned_tensor
 from vllm.v1.attention.backend import (
     AttentionBackend,
     AttentionCGSupport,
@@ -217,7 +217,7 @@ class FlashInferMLASparseMetadataBuilder(
         # Zero-fill for cudagraphs
         self.req_id_per_token_buffer.fill_(0)
         self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_(
-            torch.from_numpy(req_id_per_token), non_blocking=True
+            np_to_pinned_tensor(req_id_per_token), non_blocking=True
         )
         req_id_per_token_tensor = self.req_id_per_token_buffer[:num_tokens]
 

vllm/v1/attention/backends/mla/flashmla_sparse.py

@@ -16,7 +16,7 @@ from vllm.model_executor.layers.attention.mla_attention import (
 from vllm.platforms import current_platform
 from vllm.platforms.interface import DeviceCapability
 from vllm.utils.platform_utils import num_compute_units
-from vllm.utils.torch_utils import is_quantized_kv_cache
+from vllm.utils.torch_utils import is_quantized_kv_cache, np_to_pinned_tensor
 from vllm.v1.attention.backend import (
     AttentionBackend,
     AttentionCGSupport,
@@ -503,7 +503,7 @@ class FlashMLASparseMetadataBuilder(AttentionMetadataBuilder[FlashMLASparseMetad
         # Zero-fill for cudagraphs
         self.req_id_per_token_buffer.fill_(0)
         self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_(
-            torch.from_numpy(req_id_per_token), non_blocking=True
+            np_to_pinned_tensor(req_id_per_token), non_blocking=True
         )
         req_id_per_token = self.req_id_per_token_buffer[:num_tokens]
 

vllm/v1/attention/backends/utils.py

@@ -17,7 +17,7 @@ from typing_extensions import runtime_checkable
 
 from vllm.config import VllmConfig, get_layers_from_vllm_config
 from vllm.utils.math_utils import cdiv
-from vllm.utils.torch_utils import async_tensor_h2d
+from vllm.utils.torch_utils import PIN_MEMORY, async_tensor_h2d, np_to_pinned_tensor
 from vllm.v1.kv_cache_interface import KVCacheSpec, MambaSpec
 
 if TYPE_CHECKING:
@@ -364,8 +364,8 @@ def make_local_attention_virtual_batches(
     # tensor first, which recovers perf.
     # Upload the index tensors to the block_table's device up-front so that the
     # fancy indexing below doesn't implicitly force a synchronous H2D copy.
-    batch_indices_torch = torch.from_numpy(batch_indices).to(device, non_blocking=True)
-    block_indices_torch = torch.from_numpy(block_indices).to(device, non_blocking=True)
+    batch_indices_torch = async_tensor_h2d(batch_indices, device=device)
+    block_indices_torch = async_tensor_h2d(block_indices, device=device)
 
     # Save as a lambda so we can return this for update_block_table
     make_block_table = lambda block_table: block_table[
@@ -379,8 +379,8 @@ def make_local_attention_virtual_batches(
 
     return CommonAttentionMetadata(
         query_start_loc_cpu=query_start_loc_cpu,
-        query_start_loc=query_start_loc_cpu.to(device=device, non_blocking=True),
-        seq_lens=seq_lens_cpu.to(device=device, non_blocking=True),
+        query_start_loc=async_tensor_h2d(query_start_loc_cpu, device=device),
+        seq_lens=async_tensor_h2d(seq_lens_cpu, device=device),
         num_reqs=len(seq_lens_cpu),
         num_actual_tokens=common_attn_metadata.num_actual_tokens,
         max_query_len=seqlens_q_local.max(),
@@ -808,14 +808,12 @@ def create_fast_prefill_custom_backend(
 
 
 def compute_causal_conv1d_metadata(
-    query_start_loc_p_cpu: torch.Tensor,
-    *,
-    device: torch.device,
-):
+    query_start_loc_p_cpu: torch.Tensor, *, device: torch.device
+) -> tuple[dict[int, dict[str, Any]], torch.Tensor, torch.Tensor]:
     # Needed for causal_conv1d. Use the CPU query_start_loc to avoid DtoH sync.
     assert query_start_loc_p_cpu.device.type == "cpu"
     seqlens = query_start_loc_p_cpu.diff()
-    nums_dict = {}  # type: ignore
+    nums_dict: dict[int, dict[str, Any]] = {}
     batch_ptr = None
     token_chunk_offset_ptr = None
     for BLOCK_M in [8]:  # cover all BLOCK_M values
@@ -823,7 +821,7 @@ def compute_causal_conv1d_metadata(
         nums_dict[BLOCK_M] = {}
         nums_dict[BLOCK_M]["nums"] = nums
         nums_dict[BLOCK_M]["tot"] = nums.sum().item()
-        mlist = torch.from_numpy(np.repeat(np.arange(len(nums)), nums))
+        mlist = np_to_pinned_tensor(np.repeat(np.arange(len(nums)), nums))
         nums_dict[BLOCK_M]["mlist"] = mlist
         mlist_len = len(nums_dict[BLOCK_M]["mlist"])
         nums_dict[BLOCK_M]["mlist_len"] = mlist_len
@@ -831,7 +829,7 @@ def compute_causal_conv1d_metadata(
         offsetlist = []  # type: ignore
         for idx, num in enumerate(nums):
             offsetlist.extend(range(num))
-        offsetlist = torch.tensor(offsetlist, dtype=torch.int32)
+        offsetlist = torch.tensor(offsetlist, dtype=torch.int32, pin_memory=PIN_MEMORY)
         nums_dict[BLOCK_M]["offsetlist"] = offsetlist
 
         if batch_ptr is None:
@@ -845,16 +843,15 @@ def compute_causal_conv1d_metadata(
         else:
             if batch_ptr.nelement() < MAX_NUM_PROGRAMS:
                 batch_ptr.resize_(MAX_NUM_PROGRAMS).fill_(PAD_SLOT_ID)
-                token_chunk_offset_ptr.resize_(  # type: ignore
-                    MAX_NUM_PROGRAMS
-                ).fill_(PAD_SLOT_ID)
+                assert token_chunk_offset_ptr is not None
+                token_chunk_offset_ptr.resize_(MAX_NUM_PROGRAMS).fill_(PAD_SLOT_ID)
 
+        assert batch_ptr is not None
         batch_ptr[0:mlist_len].copy_(mlist, non_blocking=True)
-        token_chunk_offset_ptr[  # type: ignore
-            0:mlist_len
-        ].copy_(offsetlist, non_blocking=True)
+        assert token_chunk_offset_ptr is not None
+        token_chunk_offset_ptr[0:mlist_len].copy_(offsetlist, non_blocking=True)
         nums_dict[BLOCK_M]["batch_ptr"] = batch_ptr
-        nums_dict[BLOCK_M]["token_chunk_offset_ptr"] = token_chunk_offset_ptr  # type: ignore
+        nums_dict[BLOCK_M]["token_chunk_offset_ptr"] = token_chunk_offset_ptr
 
     return nums_dict, batch_ptr, token_chunk_offset_ptr
 

vllm/v1/attention/ops/rocm_aiter_mla_sparse.py

@@ -58,7 +58,9 @@ def _indexer_k_quant_and_cache_kernel(
     slot_id = tl.load(slot_mapping_ptr + tid)
     if slot_id < 0:
         return
-    block_id = slot_id // block_size
+    # The packed KV layout makes per-block strides large
+    # enough that block_id * stride can exceed 32-bit range.
+    block_id = (slot_id // block_size).to(tl.int64)
     block_offset = slot_id % block_size
     tile_block_id = block_offset // BLOCK_TILE_SIZE
     tile_block_offset = block_offset % BLOCK_TILE_SIZE
@@ -179,7 +181,9 @@ def _cp_gather_indexer_quant_cache_kernel(
         block_table_ptr + block_table_offset, mask=valid_block_table, other=-1
     )
     valid_block = valid_block_table & (block_id >= 0) & (block_id < NUM_BLOCKS)
-    safe_block_id = tl.where(valid_block, block_id, 0)
+    # The packed KV layout makes per-block strides large
+    # enough that block_id * stride can exceed 32-bit range.
+    safe_block_id = tl.where(valid_block, block_id, 0).to(tl.int64)
     safe_block_offset = tl.where(valid_block, block_offset, 0)
     tiled_block_offset = safe_block_offset % BLOCK_TILE_SIZE
     if LAYOUT == "SHUFFLE":

vllm/v1/core/kv_cache_utils.py

@@ -938,9 +938,7 @@ def _pool_bytes_per_block(kv_cache_groups: list[KVCacheGroupSpec]) -> int:
         kv_cache_groups[0].kv_cache_spec, UniformTypeKVCacheSpecs
     ):
         return kv_cache_groups[0].kv_cache_spec.page_size_bytes
-    if all(
-        isinstance(g.kv_cache_spec, UniformTypeKVCacheSpecs) for g in kv_cache_groups
-    ):
+    if _use_packed_kv_cache_groups(kv_cache_groups):
         # buckets = {page_size: [[layer_names], [layer_names], ...]}
         buckets = _bucket_layers_by_page_size(kv_cache_groups)
         return sum(ps * len(slots) for ps, slots in buckets.items())
@@ -1218,16 +1216,29 @@ def _bucket_layers_by_page_size(
     return buckets
 
 
-def _get_kv_cache_config_deepseek_v4(
+def _use_packed_kv_cache_groups(
+    kv_cache_groups: list[KVCacheGroupSpec],
+) -> bool:
+    is_dsv4 = all(
+        isinstance(group.kv_cache_spec, UniformTypeKVCacheSpecs)
+        for group in kv_cache_groups
+    )
+    return is_dsv4 or (
+        bool(envs.VLLM_USE_PACKED_HMA_KV_CACHE) and len(kv_cache_groups) > 1
+    )
+
+
+def _get_kv_cache_config_packed(
     vllm_config: VllmConfig,
     kv_cache_groups: list[KVCacheGroupSpec],
     available_memory: int,
 ) -> tuple[int, list[KVCacheTensor]]:
-    """DeepseekV4 KV cache tensor layout planning.
+    """Plan a packed per-block KV cache tensor layout.
 
     Emit one KVCacheTensor per (slot_idx, page_size). Layers from different
     groups at the same slot share a tensor (they have independent block
-    tables so block-id namespaces never collide).
+    tables so block-id namespaces never collide). Each emitted tensor aliases
+    one physical backing allocation, with per-block data laid out contiguously.
     """
     # buckets = {page_size: [[layer_names], [layer_names], ...]}
     buckets = _bucket_layers_by_page_size(kv_cache_groups)
@@ -1255,6 +1266,9 @@ def _get_kv_cache_config_deepseek_v4(
     return num_blocks, kv_cache_tensors
 
 
+_get_kv_cache_config_deepseek_v4 = _get_kv_cache_config_packed
+
+
 def get_kv_cache_config_from_groups(
     vllm_config: VllmConfig,
     kv_cache_groups: list[KVCacheGroupSpec],
@@ -1299,13 +1313,11 @@ def get_kv_cache_config_from_groups(
             )
             for layer_name in kv_cache_groups[0].layer_names
         ]
-    elif all(
-        isinstance(group.kv_cache_spec, UniformTypeKVCacheSpecs)
-        for group in kv_cache_groups
-    ):
-        # DeepseekV4: UniformTypeKVCacheSpecs but multiple groups.
-        # Delegate to the DeepseekV4-specific allocator.
-        num_blocks, kv_cache_tensors = _get_kv_cache_config_deepseek_v4(
+    elif _use_packed_kv_cache_groups(kv_cache_groups):
+        # DeepSeek V4 keeps the existing packed layout. Other multi-group
+        # attention-only HMA layouts can opt in with
+        # VLLM_USE_PACKED_HMA_KV_CACHE=1.
+        num_blocks, kv_cache_tensors = _get_kv_cache_config_packed(
             vllm_config, kv_cache_groups, available_memory
         )
     else:

vllm/v1/core/sched/output.py

@@ -118,8 +118,8 @@ class CachedRequestData:
     # NOTE(woosuk): new_token_ids is only used for pipeline parallelism.
     # When PP is not used, new_token_ids will be empty.
     new_token_ids: list[list[int]]
-    # For requests not scheduled in the last step, propagate the token ids to the
-    # connector. Won't contain requests that were scheduled in the prior step.
+    # MRV1-only: For requests not scheduled in the last step, propagate the token ids
+    # to the connector. Won't contain requests scheduled in the prior step.
     all_token_ids: dict[str, list[int]]
     new_block_ids: list[tuple[list[int], ...] | None]
     num_computed_tokens: list[int]

vllm/v1/core/sched/scheduler.py

@@ -101,6 +101,7 @@ class Scheduler(SchedulerInterface):
         self.finished_req_ids_dict: dict[int, set[str]] | None = (
             defaultdict(set) if include_finished_set else None
         )
+        # Track requests scheduled in prior step (MRV1-only).
         self.prev_step_scheduled_req_ids: set[str] = set()
 
         # Scheduling constraints.
@@ -1010,8 +1011,8 @@ class Scheduler(SchedulerInterface):
 
         # Construct the scheduler output.
         if self.use_v2_model_runner:
-            scheduled_new_reqs = scheduled_new_reqs + scheduled_resumed_reqs
-            scheduled_resumed_reqs = []
+            scheduled_new_reqs.extend(scheduled_resumed_reqs)
+            scheduled_resumed_reqs.clear()
             new_reqs_data = [
                 NewRequestData.from_request(
                     req,
@@ -1037,9 +1038,10 @@ class Scheduler(SchedulerInterface):
                 req_to_new_blocks,
             )
 
-        # Record the request ids that were scheduled in this step.
-        self.prev_step_scheduled_req_ids.clear()
-        self.prev_step_scheduled_req_ids.update(num_scheduled_tokens.keys())
+        # Record the request ids that were scheduled in this step (MRV1-only).
+        if not self.use_v2_model_runner:
+            self.prev_step_scheduled_req_ids.clear()
+            self.prev_step_scheduled_req_ids.update(num_scheduled_tokens.keys())
 
         new_block_ids_to_zero = (
             (self.kv_cache_manager.take_new_block_ids() or None)
@@ -1252,12 +1254,11 @@ class Scheduler(SchedulerInterface):
                     req.num_computed_tokens : req.num_computed_tokens + num_tokens
                 ]
                 new_token_ids.append(token_ids)
-            scheduled_in_prev_step = req_id in self.prev_step_scheduled_req_ids
             if idx >= num_running_reqs:
-                assert not scheduled_in_prev_step
                 resumed_req_ids.add(req_id)
-            if not scheduled_in_prev_step:
-                all_token_ids[req_id] = req.all_token_ids.copy()
+            if not self.use_v2_model_runner:  # noqa: SIM102
+                if req_id not in self.prev_step_scheduled_req_ids:
+                    all_token_ids[req_id] = req.all_token_ids.copy()
             new_block_ids.append(
                 req_to_new_blocks[req_id].get_block_ids(allow_none=True)
             )

vllm/v1/engine/async_llm.py

@@ -1109,9 +1109,3 @@ class AsyncLLM(EngineClient):
     async def finish_weight_update(self) -> None:
         """Finish the current weight update."""
         await self.collective_rpc("finish_weight_update")
-        # Invalidate cached state computed with the old weights so it isn't
-        # reused for subsequent requests:
-        # - prefix cache: KV blocks computed with the old weights
-        # - encoder cache: multimodal embeddings keyed only by mm_hash
-        await self.reset_prefix_cache()
-        await self.reset_encoder_cache()

vllm/v1/engine/core.py

@@ -74,7 +74,7 @@ from vllm.v1.engine.utils import (
     EngineHandshakeMetadata,
     EngineZmqAddresses,
     SignalCallback,
-    get_device_indices,
+    get_physical_gpu_ids_for_local_dp_rank,
 )
 from vllm.v1.executor import Executor
 from vllm.v1.kv_cache_interface import KVCacheConfig, get_kv_cache_spec_kind
@@ -2175,23 +2175,30 @@ class EngineCoreActorMixin:
             pass
         else:
             device_control_env_var = current_platform.device_control_env_var
-            self._set_cuda_visible_devices(
+            self._set_assigned_physical_gpu_ids(
                 vllm_config, local_dp_rank, device_control_env_var
             )
 
-    def _set_cuda_visible_devices(
-        self, vllm_config: VllmConfig, local_dp_rank: int, device_control_env_var: str
+    def _set_assigned_physical_gpu_ids(
+        self,
+        vllm_config: VllmConfig,
+        local_dp_rank: int,
+        device_control_env_var: str,
     ):
         world_size = vllm_config.parallel_config.world_size
-        # Set CUDA_VISIBLE_DEVICES or equivalent.
         try:
-            value = get_device_indices(
-                device_control_env_var, local_dp_rank, world_size
+            physical_gpu_ids = get_physical_gpu_ids_for_local_dp_rank(
+                device_control_env_var,
+                local_dp_rank,
+                world_size,
+                user_assigned_gpu_ids=(
+                    vllm_config.parallel_config.assigned_physical_gpu_ids
+                ),
             )
-            os.environ[device_control_env_var] = value
+            vllm_config.parallel_config.assigned_physical_gpu_ids = physical_gpu_ids
         except IndexError as e:
             raise Exception(
-                f"Error setting {device_control_env_var}: "
+                f"Error computing assigned_physical_gpu_ids: "
                 f"local range: [{local_dp_rank * world_size}, "
                 f"{(local_dp_rank + 1) * world_size}) "
                 f'base value: "{os.getenv(device_control_env_var)}"'

vllm/v1/engine/utils.py

@@ -12,7 +12,6 @@ from multiprocessing import Process, connection
 from multiprocessing.process import BaseProcess
 from multiprocessing.queues import Queue
 from typing import TYPE_CHECKING, cast
-from unittest.mock import patch
 
 import msgspec
 import zmq
@@ -175,38 +174,38 @@ class CoreEngineProcManager:
         self.manager_stopped = threading.Event()
         self.failed_proc_name: str | None = None
 
+        # All ranks share this config object: capture the user-provided
+        # --device-ids list before the per-rank shard overwrites it. Mutating
+        # the config before each proc.start() works because the spawn method
+        # pickles process args at start() time, sequentially per rank.
+        user_assigned_gpu_ids = vllm_config.parallel_config.assigned_physical_gpu_ids
         try:
             for proc, local_dp_rank in zip(self.processes, local_dp_ranks):
-                # Adjust device control in DP for platforms that cannot rely
-                # on torch.accelerator.set_device_index(), and for Ray launchers.
-                device_control_context: contextlib.AbstractContextManager[None] = (
-                    contextlib.nullcontext()
-                )
+                # Populate the logical-to-physical GPU mapping in DP for
+                # platforms that cannot rely on
+                # torch.accelerator.set_device_index(), and for Ray.
                 needs_device_env_isolation = not (
                     current_platform.is_cuda_alike() or current_platform.is_xpu()
                 )
                 if is_dp and (
                     needs_device_env_isolation or vllm_config.parallel_config.use_ray
                 ):
-                    device_control_context = set_device_control_env_var(
-                        vllm_config, local_dp_rank
+                    set_assigned_physical_gpu_ids_for_dp_rank(
+                        vllm_config, local_dp_rank, user_assigned_gpu_ids
                     )
 
-                with (
-                    device_control_context,
-                    numa_utils.configure_subprocess(
-                        # EngineCore itself does not have a TP/PP-local rank.
-                        # When DP is enabled, set_device_control_env_var()
-                        # narrows visible devices to this DP shard first, so
-                        # local_rank=0 means "the first local GPU in this
-                        # shard". The actual TP/PP worker processes spawned by
-                        # the executor are bound separately with their own
-                        # local_rank values.
-                        vllm_config,
-                        local_rank=0,
-                        dp_local_rank=local_dp_rank,
-                        process_kind="EngineCore",
-                    ),
+                with numa_utils.configure_subprocess(
+                    # EngineCore itself does not have a TP/PP-local rank.
+                    # When DP is enabled, set_assigned_physical_gpu_ids_for_dp_rank()
+                    # populates the logical-to-physical mapping for this DP
+                    # shard, so local_rank=0 means "the first local GPU in
+                    # this shard". The actual TP/PP worker processes spawned
+                    # by the executor are bound separately with their own
+                    # local_rank values.
+                    vllm_config,
+                    local_rank=0,
+                    dp_local_rank=local_dp_rank,
+                    process_kind="EngineCore",
                 ):
                     proc.start()
         finally:
@@ -281,55 +280,79 @@ class SignalCallback:
         self._event.set()
 
 
-@contextlib.contextmanager
-def set_device_control_env_var(
-    vllm_config: VllmConfig, local_dp_rank: int
-) -> Iterator[None]:
+def set_assigned_physical_gpu_ids_for_dp_rank(
+    vllm_config: VllmConfig,
+    local_dp_rank: int,
+    user_assigned_gpu_ids: list[int] | None = None,
+) -> None:
     """
-    Temporarily set CUDA_VISIBLE_DEVICES or equivalent
-    for engine subprocess.
+    Populate assigned_physical_gpu_ids on the config for the given DP rank.
+
+    user_assigned_gpu_ids is the full (un-sharded) --device-ids list, if the
+    user provided one; this DP rank's shard is sliced from it. It is passed
+    explicitly rather than read from the config because callers may reuse
+    one config object across DP ranks, overwriting the field each time.
     """
     world_size = vllm_config.parallel_config.world_size
     local_world_size = vllm_config.parallel_config.local_world_size
     evar = current_platform.device_control_env_var
 
-    value = get_device_indices(evar, local_dp_rank, world_size, local_world_size)
-    with patch.dict(os.environ, values=((evar, value),)):
-        yield
+    physical_gpu_ids = get_physical_gpu_ids_for_local_dp_rank(
+        evar,
+        local_dp_rank,
+        world_size,
+        local_world_size,
+        user_assigned_gpu_ids=user_assigned_gpu_ids,
+    )
+    vllm_config.parallel_config.assigned_physical_gpu_ids = physical_gpu_ids
 
 
-def get_device_indices(
+def get_physical_gpu_ids_for_local_dp_rank(
     device_control_env_var: str,
     local_dp_rank: int,
     world_size: int,
     local_world_size: int | None = None,
-):
+    user_assigned_gpu_ids: list[int] | None = None,
+) -> list[int]:
     """
-    Returns a comma-separated string of device indices for the specified
+    Returns list of physical GPU IDs for the specified
     data parallel rank.
 
     For example, if world_size=2 and local_dp_rank=1, and there are 4 devices,
-    this will select devices 2 and 3 for local_dp_rank=1.
+    this will return [2, 3] for local_dp_rank=1.
+
+    If user_assigned_gpu_ids is provided (e.g. from --device-ids), this DP
+    rank's shard is sliced from it instead of being derived from the
+    device-control env var.
     """
     if local_world_size is None:
         local_world_size = world_size
+    if user_assigned_gpu_ids is not None:
+        start = local_dp_rank * world_size
+        stop = start + local_world_size
+        if stop > len(user_assigned_gpu_ids):
+            raise ValueError(
+                f"--device-ids provides {len(user_assigned_gpu_ids)} devices, "
+                f"but DP rank {local_dp_rank} needs devices [{start}, {stop})"
+            )
+        return user_assigned_gpu_ids[start:stop]
     try:
-        value = ",".join(
-            str(current_platform.device_id_to_physical_device_id(i))
+        return [
+            current_platform.device_id_to_physical_device_id(i)
             for i in range(
                 local_dp_rank * world_size,
                 local_dp_rank * world_size + local_world_size,
             )
-        )
+        ]
     except IndexError as e:
         raise Exception(
-            f"Error setting {device_control_env_var}: "
+            f"Error computing device indices for "
+            f"{device_control_env_var}: "
             f"local range: [{local_dp_rank * world_size}, "
             f"{(local_dp_rank + 1) * world_size}) "
             "base value: "
             f'"{os.getenv(device_control_env_var)}"'
         ) from e
-    return value
 
 
 def _apply_dp_identity_suffix(dp_vllm_config, dp_rank: int) -> None:
@@ -453,11 +476,11 @@ class CoreEngineActorManager:
             # https://github.com/ray-project/ray/blob/master/python/ray/_private/accelerators/intel_gpu.py#L56 # noqa: E501
             if current_platform.is_xpu():
                 device_evar = current_platform.device_control_env_var
-                device_indices = get_device_indices(
+                physical_gpu_ids = get_physical_gpu_ids_for_local_dp_rank(
                     device_evar, local_index, world_size
                 )
                 actor_env_vars = self.env_vars_dict.copy()
-                actor_env_vars[device_evar] = device_indices
+                actor_env_vars[device_evar] = ",".join(str(d) for d in physical_gpu_ids)
                 runtime_env = RuntimeEnv(env_vars=actor_env_vars)
 
             actor = (

vllm/v1/executor/multiproc_executor.py

@@ -826,6 +826,16 @@ class WorkerProc:
         signal.signal(signal.SIGTERM, signal_handler)
         signal.signal(signal.SIGINT, signal_handler)
 
+        # Publish the logical-to-physical mapping early so topology helpers
+        # work before init_device (needed by set_worker_net_device below).
+        assigned_physical_gpu_ids = kwargs[
+            "vllm_config"
+        ].parallel_config.assigned_physical_gpu_ids
+        if assigned_physical_gpu_ids is not None:
+            from vllm.platforms.interface import set_assigned_physical_gpu_ids
+
+            set_assigned_physical_gpu_ids(assigned_physical_gpu_ids)
+
         # Set net device env vars for the worker if VLLM_GPU_NIC_PCIE_MAPPING is set
         set_worker_net_device(kwargs.get("local_rank", 0), kwargs["vllm_config"])
 

vllm/v1/executor/ray_executor.py

@@ -258,30 +258,35 @@ class RayDistributedExecutor(Executor):
         }
         self.collective_rpc("adjust_rank", args=(rerank_mapping,))
 
-        # Get the set of GPU IDs used on each node.
-        worker_node_and_gpu_ids = []
+        # Get the set of physical GPU IDs used on each node.
+        worker_node_and_physical_gpu_ids = []
         for worker in [self.driver_dummy_worker] + self.workers:
             if worker is None:
                 # driver_dummy_worker can be None when using ray spmd worker.
                 continue
-            worker_node_and_gpu_ids.append(
-                ray.get(worker.get_node_and_gpu_ids.remote())  # type: ignore[attr-defined]
+            worker_node_and_physical_gpu_ids.append(
+                ray.get(worker.get_node_and_physical_gpu_ids.remote())  # type: ignore[attr-defined]
             )
 
         node_workers = defaultdict(list)  # node id -> list of worker ranks
-        node_gpus = defaultdict(list)  # node id -> list of gpu ids
+        node_physical_gpu_ids = defaultdict(list)  # node id -> physical GPU IDs
 
-        for i, (node_id, gpu_ids) in enumerate(worker_node_and_gpu_ids):
+        for i, (node_id, physical_gpu_ids) in enumerate(
+            worker_node_and_physical_gpu_ids
+        ):
             node_workers[node_id].append(i)
-            # `gpu_ids` can be a list of strings or integers.
+            # `physical_gpu_ids` can be a list of strings or integers.
             # convert them to integers for consistency.
-            # NOTE: gpu_ids can be larger than 9 (e.g. 16 GPUs),
+            # NOTE: physical GPU IDs can be larger than 9 (e.g. 16 GPUs),
             # string sorting is not sufficient.
             # see https://github.com/vllm-project/vllm/issues/5590
-            gpu_ids = [int(x) for x in gpu_ids]
-            node_gpus[node_id].extend(gpu_ids)
-        for node_id, gpu_ids in node_gpus.items():
-            node_gpus[node_id] = sorted(gpu_ids)
+            physical_gpu_ids = [
+                current_platform.device_control_id_to_physical_device_id(str(x))
+                for x in physical_gpu_ids
+            ]
+            node_physical_gpu_ids[node_id].extend(physical_gpu_ids)
+        for node_id, physical_gpu_ids in node_physical_gpu_ids.items():
+            node_physical_gpu_ids[node_id] = sorted(physical_gpu_ids)
 
         all_ips = set(worker_ips + [driver_ip])
         n_ips = len(all_ips)
@@ -297,23 +302,8 @@ class RayDistributedExecutor(Executor):
                 " each node."
             )
 
-        # Set environment variables for the driver and workers.
-        # We set CUDA_VISIBLE_DEVICES to ALL GPUs on the node for each worker.
-        # This is needed because:
-        # 1. Ray's compiled DAG needs to find the allocated GPU in
-        #    CUDA_VISIBLE_DEVICES.
-        # 2. vLLM's communication layer (NCCL, CustomAllreduce) needs to see
-        #    all GPUs for P2P checks and communication setup. Though if it was
-        #    just this reason, we could have also just kept the visible devices
-        #    unset.
-        # Each worker will use local_rank to index into the visible devices.
-        all_args_to_update_environment_variables = [
-            {
-                current_platform.device_control_env_var: ",".join(
-                    map(str, node_gpus[node_id])
-                ),
-            }
-            for (node_id, _) in worker_node_and_gpu_ids
+        all_args_to_update_environment_variables: list[dict[str, str]] = [
+            {} for _ in worker_node_and_physical_gpu_ids
         ]
 
         # Environment variables to copy from driver to workers
@@ -336,7 +326,7 @@ class RayDistributedExecutor(Executor):
             "update_environment_variables", args=(self._get_env_vars_to_be_updated(),)
         )
 
-        if len(node_gpus) == 1:
+        if len(node_physical_gpu_ids) == 1:
             # in single node case, we don't need to get the IP address.
             # the loopback address is sufficient
             # NOTE: a node may have several IP addresses, one for each
@@ -352,10 +342,11 @@ class RayDistributedExecutor(Executor):
 
         # Initialize the actual workers inside worker wrapper.
         all_kwargs = []
-        for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids):
+        for rank, (node_id, _) in enumerate(worker_node_and_physical_gpu_ids):
             local_rank = node_workers[node_id].index(rank)
             kwargs = dict(
                 vllm_config=self.vllm_config,
+                assigned_physical_gpu_ids=sorted(node_physical_gpu_ids[node_id]),
                 local_rank=local_rank,
                 rank=rank,
                 distributed_init_method=distributed_init_method,

vllm/v1/executor/ray_executor_v2.py

@@ -79,24 +79,25 @@ class RayWorkerProc(WorkerProc):
     1. __init__: lightweight setup, stores init args (no device/model init)
     2. initialize_worker: called after GPU IDs are discovered, completes
        the full WorkerProc initialization with the correct local_rank and
-       CUDA_VISIBLE_DEVICES.
+       logical-to-physical GPU mapping.
 
-    CUDA_VISIBLE_DEVICES setup flow:
+    GPU assignment flow:
 
     1. RayExecutorV2 enables RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES so Ray does
        not set CUDA_VISIBLE_DEVICES on RayWorkerProc actors at creation time.
     2. Each actor is scheduled with a placement group and bundle index; Ray resolves
        the physical GPU ID for that bundle at placement time.
-    3. After placement, the worker discovers that GPU ID and sets
-       CUDA_VISIBLE_DEVICES before finishing WorkerProc initialization.
+    3. After placement, the executor discovers each worker's GPU ID and passes the
+       node's logical-to-physical mapping (assigned_physical_gpu_ids) to
+       initialize_worker(); CUDA_VISIBLE_DEVICES is never modified.
 
-    There is no workaround for this unset-and-reset sequence when the placement group
-    is externally managed: scheduling must complete before CUDA_VISIBLE_DEVICES can
-    match the GPU tied to the worker's bundle.
+    Scheduling must complete before the mapping is known when the placement
+    group is externally managed: only then is the GPU tied to the worker's
+    bundle resolved.
 
     This sequence allows multiple vLLM instances to coexist on the same node:
     each instance is unaware which physical devices others hold, and the
-    externally managed placement group avoids CUDA_VISIBLE_DEVICES conflicts
+    externally managed placement group avoids device assignment conflicts
     by binding workers to specific placement group bundles.
     """
 
@@ -120,28 +121,33 @@ class RayWorkerProc(WorkerProc):
             is_driver_worker=is_driver_worker,
         )
 
-    def get_node_and_gpu_ids(self) -> tuple[str, list[int]]:
-        """Return (node_id, gpu_ids) assigned to this actor by Ray."""
+    def get_node_and_physical_gpu_ids(self) -> tuple[str, list[int]]:
+        """Return (node_id, physical_gpu_ids) assigned to this actor by Ray."""
         node_id = ray.get_runtime_context().get_node_id()
         device_key = current_platform.ray_device_key
         if not device_key:
             raise RuntimeError(
                 f"current platform {current_platform.device_name} does not support ray."
             )
-        gpu_ids = ray.get_runtime_context().get_accelerator_ids()[device_key]
-        return node_id, [int(x) for x in gpu_ids]
+        physical_gpu_ids = ray.get_runtime_context().get_accelerator_ids()[device_key]
+        return node_id, [
+            current_platform.device_control_id_to_physical_device_id(str(x))
+            for x in physical_gpu_ids
+        ]
 
     def initialize_worker(
         self,
         local_rank: int,
         env_vars: dict[str, str],
         driver_env_vars: dict[str, str] | None = None,
+        assigned_physical_gpu_ids: list[int] | None = None,
     ) -> None:
         """Complete initialization after GPU assignment is known.
 
         *driver_env_vars* are applied with ``setdefault`` — they fill
         in missing vars but never overwrite node-local values.
-        *env_vars* (e.g. CUDA_VISIBLE_DEVICES) always overwrite.
+        *env_vars* always overwrite.
+        *assigned_physical_gpu_ids* maps local_rank to physical CUDA device ID.
         """
         if driver_env_vars:
             for key, value in driver_env_vars.items():
@@ -149,6 +155,13 @@ class RayWorkerProc(WorkerProc):
         for key, value in env_vars.items():
             os.environ[key] = value
 
+        if assigned_physical_gpu_ids is not None:
+            vllm_config = self._init_kwargs["vllm_config"]
+            assert isinstance(vllm_config, VllmConfig)
+            vllm_config.parallel_config.assigned_physical_gpu_ids = (
+                assigned_physical_gpu_ids
+            )
+
         self.local_rank = local_rank
         super().__init__(
             local_rank=local_rank,
@@ -365,36 +378,48 @@ class RayExecutorV2(MultiprocExecutor):
             )
             self.ray_worker_handles.append(handle)
 
-        # Step 6: Discover GPU IDs assigned to each worker via Ray runtime context.
-        worker_node_and_gpu_ids = ray.get(
-            [h.actor.get_node_and_gpu_ids.remote() for h in self.ray_worker_handles]
+        # Step 6: Discover physical GPU IDs assigned to each worker via Ray
+        # runtime context.
+        worker_node_and_physical_gpu_ids = ray.get(
+            [
+                h.actor.get_node_and_physical_gpu_ids.remote()
+                for h in self.ray_worker_handles
+            ]
         )
 
         node_workers: dict[str, list[int]] = defaultdict(list)
-        node_gpus: dict[str, list[int]] = defaultdict(list)
-        for i, (node_id, gpu_ids) in enumerate(worker_node_and_gpu_ids):
+        node_physical_gpu_ids: dict[str, list[int]] = defaultdict(list)
+        for i, (node_id, physical_gpu_ids) in enumerate(
+            worker_node_and_physical_gpu_ids
+        ):
             node_workers[node_id].append(i)
-            node_gpus[node_id].extend(gpu_ids)
-        for node_id, gpu_ids in node_gpus.items():
-            node_gpus[node_id] = sorted(gpu_ids)
+            node_physical_gpu_ids[node_id].extend(physical_gpu_ids)
+        for node_id, physical_gpu_ids in node_physical_gpu_ids.items():
+            node_physical_gpu_ids[node_id] = sorted(physical_gpu_ids)
 
-        # Step 7: Initialize workers with correct local_rank and
-        # CUDA_VISIBLE_DEVICES. Each worker sees all GPUs assigned to
-        # this executor on its node; local_rank indexes into that set.
+        # Step 7: Initialize workers with local logical ranks and the
+        # logical-to-physical GPU mapping discovered from Ray placement.
         init_worker_refs = []
-        for i, (node_id, _) in enumerate(worker_node_and_gpu_ids):
+        for i, (node_id, _) in enumerate(worker_node_and_physical_gpu_ids):
             local_rank = node_workers[node_id].index(i)
-            worker_env_vars = {
-                current_platform.device_control_env_var: ",".join(
-                    map(str, node_gpus[node_id])
-                ),
-            }
+            assigned_physical_gpu_ids = sorted(node_physical_gpu_ids[node_id])
+            worker_env_vars: dict[str, str] = {}
             self.ray_worker_handles[i].local_rank = local_rank
             init_worker_refs.append(
                 self.ray_worker_handles[i].actor.initialize_worker.remote(
-                    local_rank, worker_env_vars, self.driver_env_vars
+                    local_rank,
+                    worker_env_vars,
+                    self.driver_env_vars,
+                    assigned_physical_gpu_ids=assigned_physical_gpu_ids,
                 )
             )
+        # Also set on the executor-side config for consistency. The mapping
+        # is per-node, so only do this when all workers share one node.
+        if len(node_physical_gpu_ids) == 1:
+            node_id_0 = worker_node_and_physical_gpu_ids[0][0]
+            self.vllm_config.parallel_config.assigned_physical_gpu_ids = sorted(
+                node_physical_gpu_ids[node_id_0]
+            )
         ray.get(init_worker_refs)
 
         # Step 8: Collect response MQ handles

vllm/v1/executor/ray_utils.py

@@ -93,7 +93,7 @@ try:
         def get_node_ip(self) -> str:
             return get_ip()
 
-        def get_node_and_gpu_ids(self) -> tuple[str, list[int]]:
+        def get_node_and_physical_gpu_ids(self) -> tuple[str, list[int]]:
             node_id = ray.get_runtime_context().get_node_id()
             device_key = vllm.platforms.current_platform.ray_device_key
             if not device_key:
@@ -101,8 +101,10 @@ try:
                     "current platform %s does not support ray.",
                     vllm.platforms.current_platform.device_name,
                 )
-            gpu_ids = ray.get_runtime_context().get_accelerator_ids()[device_key]
-            return node_id, gpu_ids
+            physical_gpu_ids = ray.get_runtime_context().get_accelerator_ids()[
+                device_key
+            ]
+            return node_id, physical_gpu_ids
 
         def setup_device_if_necessary(self):
             # TODO(swang): This is needed right now because Ray CG executes

vllm/v1/kv_offload/cpu/common.py

@@ -4,7 +4,10 @@ from typing_extensions import override
 
 from vllm.v1.kv_offload.base import BlockIDsLoadStoreSpec
 
-METRIC_STORES_SKIPPED = "vllm:kv_offload_stores_skipped"
+
+class CPUOffloadingMetrics:
+    STORES_SKIPPED = "vllm:kv_offload_stores_skipped"
+    CPU_CACHE_USAGE_PERC = "vllm:kv_offload_cpu_cache_usage_perc"
 
 
 class CPULoadStoreSpec(BlockIDsLoadStoreSpec):

vllm/v1/kv_offload/cpu/gpu_worker.py

@@ -14,7 +14,7 @@ from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.triton_utils import HAS_TRITON, triton
 from vllm.utils.math_utils import cdiv
-from vllm.utils.platform_utils import is_pin_memory_available
+from vllm.utils.torch_utils import PIN_MEMORY
 from vllm.v1.kv_offload.base import (
     BlockIDsLoadStoreSpec,
     CanonicalKVCacheRef,
@@ -156,7 +156,7 @@ def pin_mmap_region(region: SharedOffloadRegion) -> None:
 def _new_descriptor_buffers(
     num_copy_ops: int,
 ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    pin = is_pin_memory_available()
+    pin = PIN_MEMORY
     # CUDA cache_kernels.cu requires int64; XPU DMA engine requires uint64.
     ptr_dtype = torch.uint64 if current_platform.is_xpu() else torch.int64
     return (
@@ -482,7 +482,7 @@ class CpuGpuOffloadingHandlers:
         num_cpu_blocks: int,
         mmap_region: SharedOffloadRegion | None = None,
     ):
-        pin_memory = is_pin_memory_available()
+        pin_memory = PIN_MEMORY
         logger.info("Allocating %d CPU tensors...", len(kv_caches.tensors))
         self._mmap_region = mmap_region
         if mmap_region is not None and pin_memory:

vllm/v1/kv_offload/cpu/manager.py

@@ -18,7 +18,10 @@ from vllm.v1.kv_offload.base import (
     ReqContext,
     RequestOffloadingContext,
 )
-from vllm.v1.kv_offload.cpu.common import METRIC_STORES_SKIPPED, CPULoadStoreSpec
+from vllm.v1.kv_offload.cpu.common import (
+    CPULoadStoreSpec,
+    CPUOffloadingMetrics,
+)
 from vllm.v1.kv_offload.cpu.policies.arc import ARCCachePolicy
 from vllm.v1.kv_offload.cpu.policies.base import BlockStatus, CachePolicy
 from vllm.v1.kv_offload.cpu.policies.lru import LRUCachePolicy
@@ -282,13 +285,21 @@ class CPUOffloadingManager(OffloadingManager):
             self.events.clear()
 
     def get_stats(self) -> OffloadingConnectorStats | None:
-        if self.store_threshold < 2:
-            return None
-
         stats = OffloadingConnectorStats()
-        stats.increase_counter(
-            METRIC_STORES_SKIPPED,
-            self.stores_skipped_in_current_batch,
+
+        # Compute cache usage.
+        num_used = (
+            self._num_allocated_blocks
+            - len(self._free_list)
+            - self._num_evictable_cache_blocks
         )
-        self.stores_skipped_in_current_batch = 0
+        usage = num_used / self._num_blocks if self._num_blocks > 0 else 0.0
+        stats.set_gauge(CPUOffloadingMetrics.CPU_CACHE_USAGE_PERC, usage)
+
+        if self.store_threshold >= 2:
+            stats.increase_counter(
+                CPUOffloadingMetrics.STORES_SKIPPED,
+                self.stores_skipped_in_current_batch,
+            )
+            self.stores_skipped_in_current_batch = 0
         return stats

vllm/v1/kv_offload/cpu/spec.py

@@ -14,11 +14,15 @@ from vllm.v1.kv_offload.base import (
     GPULoadStoreSpec,
     LoadStoreSpec,
     OffloadingCounterMetadata,
+    OffloadingGaugeMetadata,
     OffloadingManager,
     OffloadingMetricMetadata,
     OffloadingSpec,
 )
-from vllm.v1.kv_offload.cpu.common import METRIC_STORES_SKIPPED, CPULoadStoreSpec
+from vllm.v1.kv_offload.cpu.common import (
+    CPULoadStoreSpec,
+    CPUOffloadingMetrics,
+)
 from vllm.v1.kv_offload.cpu.gpu_worker import CpuGpuOffloadingHandlers
 from vllm.v1.kv_offload.cpu.manager import CPUOffloadingManager
 from vllm.v1.kv_offload.worker.worker import OffloadingHandler
@@ -31,17 +35,27 @@ class CPUOffloadingSpec(OffloadingSpec):
     def build_metric_definitions(
         cls, extra_config: dict[str, Any]
     ) -> dict[str, OffloadingMetricMetadata]:
-        store_threshold = int(extra_config.get("store_threshold", 0))
-        if store_threshold < 2:
-            return {}
-        return {
-            METRIC_STORES_SKIPPED: OffloadingCounterMetadata(
+        definitions: dict[str, OffloadingMetricMetadata] = {
+            CPUOffloadingMetrics.CPU_CACHE_USAGE_PERC: OffloadingGaugeMetadata(
                 documentation=(
-                    "Number of KV offload stores skipped because the reuse "
-                    "threshold was not reached."
+                    "Fraction of CPU KV-cache space currently pinned by active "
+                    "transfers (0.0 = idle, 1.0 = saturated). Sustained high "
+                    "values indicate transfers (stores or promotions) may be "
+                    "dropped due to insufficient capacity."
                 ),
             )
         }
+        store_threshold = int(extra_config.get("store_threshold", 0))
+        if store_threshold >= 2:
+            definitions[CPUOffloadingMetrics.STORES_SKIPPED] = (
+                OffloadingCounterMetadata(
+                    documentation=(
+                        "Number of KV offload stores skipped because the reuse "
+                        "threshold was not reached."
+                    ),
+                )
+            )
+        return definitions
 
     def __init__(self, vllm_config: VllmConfig, kv_cache_config: KVCacheConfig):
         super().__init__(vllm_config, kv_cache_config)
@@ -58,7 +72,15 @@ class CPUOffloadingSpec(OffloadingSpec):
         self.cpu_page_size_per_worker = 0
         assert kv_cache_config is not None
         if kv_cache_config.num_blocks > 0 and world_size > 0:
-            total_gpu_kv_bytes = sum(t.size for t in kv_cache_config.kv_cache_tensors)
+            is_packed = any(t.block_stride for t in kv_cache_config.kv_cache_tensors)
+            assert not is_packed or all(
+                t.block_stride for t in kv_cache_config.kv_cache_tensors
+            )
+            total_gpu_kv_bytes = (
+                kv_cache_config.kv_cache_tensors[0].size
+                if is_packed
+                else sum(t.size for t in kv_cache_config.kv_cache_tensors)
+            )
             kv_bytes_per_block = (
                 total_gpu_kv_bytes // kv_cache_config.num_blocks
             ) * world_size

vllm/v1/pool/metadata.py

@@ -7,9 +7,7 @@ import torch
 
 from vllm.pooling_params import PoolingParams
 from vllm.tasks import PoolingTask
-from vllm.utils.platform_utils import is_pin_memory_available
-
-pin_memory = is_pin_memory_available()
+from vllm.utils.torch_utils import PIN_MEMORY
 
 
 @dataclass
@@ -134,7 +132,7 @@ class PoolingMetadata:
         num_scheduled_tokens_cpu = torch.from_numpy(num_scheduled_tokens_np)
         if query_start_loc_gpu is None:
             cumsum = torch.zeros(
-                n_seq + 1, dtype=torch.int64, pin_memory=pin_memory, device="cpu"
+                n_seq + 1, dtype=torch.int64, pin_memory=PIN_MEMORY, device="cpu"
             )
             torch.cumsum(num_scheduled_tokens_cpu, dim=0, out=cumsum[1:])
             cumsum = cumsum.to(device, non_blocking=True)

vllm/v1/sample/logits_processor/builtin.py

@@ -7,6 +7,7 @@ import numpy as np
 import torch
 
 from vllm import SamplingParams
+from vllm.utils.torch_utils import async_tensor_h2d
 from vllm.v1.sample.logits_processor.interface import (
     BatchUpdate,
     LogitsProcessor,
@@ -118,7 +119,6 @@ class MinPLogitsProcessor(LogitsProcessor):
 class LogitBiasLogitsProcessor(LogitsProcessor):
     def __init__(self, _, device: torch.device, is_pin_memory: bool):
         self.device = device
-        self.pin_memory = is_pin_memory
         self.biases: dict[int, dict[int, float]] = {}
 
         self.bias_tensor: torch.Tensor = torch.tensor(())
@@ -154,9 +154,7 @@ class LogitBiasLogitsProcessor(LogitsProcessor):
             )
 
     def _device_tensor(self, data: list, dtype: torch.dtype) -> torch.Tensor:
-        return torch.tensor(
-            data, device="cpu", dtype=dtype, pin_memory=self.pin_memory
-        ).to(device=self.device, non_blocking=True)
+        return async_tensor_h2d(data, device=self.device, dtype=dtype)
 
     def apply(self, logits: torch.Tensor) -> torch.Tensor:
         if self.biases:
@@ -170,7 +168,6 @@ class MinTokensLogitsProcessor(LogitsProcessor):
     ):
         # index -> (min_toks, output_token_ids, stop_token_ids)
         self.device = device
-        self.pin_memory = is_pin_memory
         self.min_toks: dict[int, tuple[int, Sequence[int], set[int]]] = {}
 
         # (req_idx_tensor,eos_tok_id_tensor)
@@ -227,9 +224,7 @@ class MinTokensLogitsProcessor(LogitsProcessor):
             )
 
     def _device_tensor(self, data: list, dtype: torch.dtype) -> torch.Tensor:
-        return torch.tensor(
-            data, device="cpu", dtype=dtype, pin_memory=self.pin_memory
-        ).to(device=self.device, non_blocking=True)
+        return async_tensor_h2d(data, device=self.device, dtype=dtype)
 
     def apply(self, logits: torch.Tensor) -> torch.Tensor:
         if self.min_toks:
@@ -283,8 +278,8 @@ class MinTokensLogitsProcessor(LogitsProcessor):
             toks_arr = np.concatenate(all_toks)
             # (row_indices, token_indices) for index_put_ to set -inf.
             logits_slice = (
-                torch.from_numpy(rows_arr).to(self.device, non_blocking=True),
-                torch.from_numpy(toks_arr).to(self.device, non_blocking=True),
+                async_tensor_h2d(rows_arr, device=self.device),
+                async_tensor_h2d(toks_arr, device=self.device),
             )
             logits.index_put_(logits_slice, self.neg_inf_tensor)
 

vllm/v1/sample/ops/penalties.py

@@ -4,8 +4,7 @@
 import torch
 
 from vllm.model_executor.layers.utils import apply_penalties
-from vllm.utils.platform_utils import is_pin_memory_available
-from vllm.utils.torch_utils import make_tensor_with_pad
+from vllm.utils.torch_utils import PIN_MEMORY, make_tensor_with_pad
 
 
 def apply_all_penalties(
@@ -52,6 +51,6 @@ def _convert_to_tensors(
         pad=vocab_size,
         device="cpu",
         dtype=torch.int64,
-        pin_memory=is_pin_memory_available(),
+        pin_memory=PIN_MEMORY,
     )
     return output_tokens_tensor.to(device, non_blocking=True)

vllm/v1/sample/sampler.py

@@ -6,7 +6,7 @@ import torch
 import torch.nn as nn
 
 from vllm.config.model import LogprobsMode
-from vllm.utils.platform_utils import is_pin_memory_available
+from vllm.utils.torch_utils import PIN_MEMORY
 from vllm.v1.outputs import LogprobsTensors, SamplerOutput
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.sample.ops.bad_words import apply_bad_words
@@ -65,7 +65,7 @@ class Sampler(nn.Module):
     ):
         super().__init__()
         self.topk_topp_sampler = TopKTopPSampler(logprobs_mode, use_fp64_gumbel)
-        self.pin_memory = is_pin_memory_available()
+        self.pin_memory = PIN_MEMORY
         self.logprobs_mode = logprobs_mode
         self.use_fp64_gumbel = use_fp64_gumbel
 

vllm/v1/sample/thinking_budget_state.py

@@ -7,7 +7,7 @@ from typing import TYPE_CHECKING, Any
 import torch
 
 from vllm.platforms import current_platform
-from vllm.utils.torch_utils import async_tensor_h2d
+from vllm.utils.torch_utils import PIN_MEMORY, async_tensor_h2d
 from vllm.v1.sample.logits_processor.interface import (
     BatchUpdate,
     MoveDirectionality,
@@ -22,12 +22,11 @@ def maybe_create_thinking_budget_state_holder(
     max_num_seqs: int,
     num_spec_tokens: int,
     device: torch.device,
-    is_pin_memory: bool,
 ) -> "ThinkingBudgetStateHolder | None":
     if reasoning_config is None:
         return None
     return ThinkingBudgetStateHolder(
-        reasoning_config, max_num_seqs, num_spec_tokens, device, is_pin_memory
+        reasoning_config, max_num_seqs, num_spec_tokens, device, PIN_MEMORY
     )