test memory adjustment

Signed-off-by: Nick Hill <nickhill123@gmail.com>

tests/evals/gsm8k/configs/humming/Qwen3.5-35B-A3B-FP8-humming-act-fp8.yaml

@@ -8,5 +8,6 @@ server_args: >-
   --max-model-len 8192
   --moe-backend humming
   --linear-backend humming
+  --gpu-memory-utilization 0.85
 env:
   VLLM_HUMMING_INPUT_QUANT_CONFIG: '{"dtype":"float8e4m3"}'

tests/evals/gsm8k/configs/humming/Qwen3.5-35B-A3B-FP8-humming.yaml

@@ -8,3 +8,4 @@ server_args: >-
   --max-model-len 8192
   --moe-backend humming
   --linear-backend humming
+  --gpu-memory-utilization 0.85

tests/evals/gsm8k/configs/humming/Qwen3.5-35B-A3B-experts-int8-humming-act-int8.yaml

@@ -8,5 +8,6 @@ server_args: >-
   --max-model-len 8192
   --moe-backend humming
   --quantization experts_int8
+  --gpu-memory-utilization 0.85
 env:
   VLLM_HUMMING_INPUT_QUANT_CONFIG: '{"dtype":"int8"}'

tests/evals/gsm8k/configs/humming/Qwen3.5-35B-A3B-experts-int8-humming.yaml

@@ -8,3 +8,4 @@ server_args: >-
   --max-model-len 8192
   --moe-backend humming
   --quantization experts_int8
+  --gpu-memory-utilization 0.85

tests/evals/gsm8k/configs/humming/Qwen3.6-35B-A3B-NVFP4-humming.yaml

@@ -8,3 +8,4 @@ server_args: >-
   --max-model-len 8192
   --moe-backend humming
   --linear-backend humming
+  --gpu-memory-utilization 0.85

tests/test_config.py

@@ -103,7 +103,7 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
     [
         (
             SimpleNamespace(
-                model="Qwen/Qwen3-1.7B-Base",
+                model="Qwen/Qwen3-32B",
                 architectures=["Qwen3ForCausalLM"],
                 runner_type="generate",
                 is_moe=False,
@@ -113,8 +113,8 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
         ),
         (
             SimpleNamespace(
-                model="Qwen/Qwen3-32B",
-                architectures=["Qwen3ForCausalLM"],
+                model="Qwen/Qwen2-7B-Instruct",
+                architectures=["Qwen2ForCausalLM"],
                 runner_type="generate",
                 is_moe=False,
                 is_quantized=False,
@@ -181,6 +181,16 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
             ),
             True,
         ),
+        (
+            SimpleNamespace(
+                model="deepseek-ai/DeepSeek-V3",
+                architectures=["DeepseekV3ForCausalLM"],
+                runner_type="generate",
+                is_moe=True,
+                is_quantized=False,
+            ),
+            True,
+        ),
         (
             SimpleNamespace(
                 model="Qwen/Qwen1.5-MoE-A2.7B",
@@ -239,7 +249,7 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
                 is_moe=True,
                 is_quantized=False,
             ),
-            False,
+            True,
         ),
         (
             SimpleNamespace(
@@ -260,7 +270,7 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
                 is_quantized=False,
                 is_hybrid=True,
             ),
-            False,
+            True,
         ),
         (
             SimpleNamespace(
@@ -271,7 +281,7 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
                 is_quantized=False,
                 is_attention_free=True,
             ),
-            False,
+            True,
         ),
         (
             SimpleNamespace(

tests/v1/e2e/general/test_mamba_prefix_cache.py

@@ -840,6 +840,10 @@ def get_mamba_prefix_cache_step_configs(
 def _run_mamba_prefix_cache_mrv1(
     monkeypatch: pytest.MonkeyPatch, async_scheduling: bool
 ):
+    # This test patches the V1 model runner, so pin V1 explicitly: MoE/hybrid
+    # models like Qwen3-Next now default to the V2 runner.
+    monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "0")
+    envs.disable_envs_cache()
     global async_scheduling_mode
     async_scheduling_mode = async_scheduling
     run_ref_mamba_state_in_subprocess()

tests/v1/e2e/spec_decode/test_spec_decode.py

@@ -562,6 +562,7 @@ def _run_eagle_correctness(
 
 
 @single_gpu_only
+@large_gpu_mark(min_gb=24)
 @pytest.mark.skipif(
     current_platform.is_device_capability_family(100),
     reason="DeepSeek head_dim=192 not supported on SM100/SM110 (Blackwell)",

tests/v1/e2e/test_hybrid_chunked_prefill.py

@@ -79,6 +79,10 @@ def test_mtp_speculative_mixed_batch_short_prefill(
         enable_chunked_prefill=True,
         enable_prefix_caching=enable_prefix_caching,
         mamba_cache_mode="align" if enable_prefix_caching else "none",
+        # MRV2 warmup materializes the worst-case spec-decode logits all-gather
+        # (num_reqs * (1 + num_speculative_tokens) rows), which memory profiling
+        # does not yet account for. Leave headroom so warmup doesn't OOM.
+        gpu_memory_utilization=0.85,
     ) as llm:
         sampling_params = SamplingParams(
             temperature=0.0,

vllm/config/vllm.py

@@ -65,17 +65,6 @@ else:
 
 logger = init_logger(__name__)
 
-DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES = frozenset(
-    {
-        "DeepseekV2ForCausalLM",
-        "GraniteMoeForCausalLM",
-        "InklingForCausalLM",
-        "InklingForConditionalGeneration",
-        "LongcatFlashNgramForCausalLM",
-        "Qwen2MoeForCausalLM",
-    }
-)
-
 
 class OptimizationLevel(IntEnum):
     """Optimization level enum."""
@@ -604,22 +593,7 @@ class VllmConfig:
 
     def _is_default_v2_model_runner_model(self) -> bool:
         model_config = self.model_config
-        if model_config is None:
-            return False
-
-        if model_config.runner_type != "generate":
-            return False
-
-        if getattr(model_config, "is_hybrid", False):
-            return False
-
-        if getattr(model_config, "is_attention_free", False):
-            return False
-        architectures = getattr(model_config, "architectures", [])
-        return (
-            any(arch in DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES for arch in architectures)
-            or not model_config.is_moe
-        )
+        return model_config is not None and model_config.runner_type == "generate"
 
     @property
     def needs_dp_coordinator(self) -> bool:

vllm/v1/worker/gpu/model_runner.py

@@ -303,7 +303,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
 
         self.model_memory_usage = m.consumed_memory
         logger.info(
-            "Model loading took %s GiB and %.6f seconds",
+            "Model loading took %s GiB memory and %.6f seconds",
             format_gib(m.consumed_memory),
             time_after_load - time_before_load,
         )
@@ -1590,6 +1590,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         """Release GPU tensors (model weights, KV caches, workspace) so that
         memory is reclaimable when running in the same process."""
         torch.accelerator.synchronize()
+        self.cudagraph_manager = None
         if hasattr(self, "kv_caches"):
             self.kv_caches.clear()
         if hasattr(self, "attn_groups"):

vllm/v1/worker/gpu/model_states/__init__.py

@@ -27,7 +27,7 @@ def init_model_state(
 
         return EncoderDecoderModelState(vllm_config, model, encoder_cache, device)
 
-    if vllm_config.model_config.is_hybrid:
+    if vllm_config.model_config.is_hybrid or vllm_config.model_config.is_attention_free:
         from vllm.v1.worker.gpu.model_states.mamba_hybrid import MambaHybridModelState
 
         return MambaHybridModelState(vllm_config, model, encoder_cache, device)