Woosuk KwonandClaude 50cfbbca7a [MRV2] Draw 64-bit uniforms for fp32 Gumbel sampling ...

tl.rand only resolves the uniform down to 2**-31, which hard-caps the
fp32 Gumbel noise at ~21.5 and coarsely buckets the argmax-deciding
tail. Philox already generates 128 bits per call (tl.rand discards 96),
so draw 64 bits instead and convert uint64 -> fp32 directly: the tail
now resolves down to 2**-64 (noise up to ~44.4) with no fp64 arithmetic
and no measurable cost. The fp64 path is unchanged.

Co-authored-by: Claude
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>

2026-07-03 06:22:59 +00:00

..

assets

[Bugfix] Align OpenCV video metadata timeline (#47099)

2026-06-30 20:43:42 +00:00

benchmarks

[Bench][BugFix] Fix empty decoder prompt for Cohere ASR in throughput benchmark (#47135)

2026-07-01 15:42:27 +00:00

compilation

[ROCm][MLA] Fuse MLA q/kv RMSNorm + FP8 per-token quant in the FP8 attention path (#44977)

2026-07-02 13:00:41 +08:00

config

[ModelRunner V2] Enable by default for all dense models (#44443)

2026-07-02 18:48:57 +08:00

cute_utils

[GDN] Improve kkt kernel of CuteDSL prefill backend (#46346)

2026-06-29 18:34:18 -07:00

device_allocator

[Core] Make sleep-mode backend capability flags communicator-agnostic (#47243)

2026-07-01 07:17:44 +00:00

distributed

Weight sync refactor + move sparse nccl engine (#44353)

2026-07-01 01:25:19 -07:00

engine

Weight sync refactor + move sparse nccl engine (#44353)

2026-07-01 01:25:19 -07:00

entrypoints

[Bugfix] Expose usage field in GenerateResponse for disaggregated serving (#42748)

2026-07-01 10:00:19 +00:00

inputs

[Render][Speculator] Add return_loss_mask to render endpoint for training data generation (#46846)

2026-06-28 00:07:33 -07:00

ir

Speed up docs build (#44635)

2026-06-05 14:51:44 +00:00

kernels

[XPU] [RMSNorm] revert weightless change on xpu (#46987)

2026-06-29 19:03:06 +00:00

logging_utils

[vLLM IR] 1/N Implement IR skeleton and rms_norm op (#33825)

2026-03-31 22:15:05 -04:00

lora

[xpu][lora]: Align LoRA implementation with Punica GPU: fix _apply_expand rank mismatch, add_inputs hardcode, and MoE EP (#45368)

2026-07-01 22:14:04 +08:00

model_executor

[Bugfix][Gemma4] Keep image bidirectional attention within the sliding window (#47217)

2026-07-02 19:57:41 -07:00

models

[BugFix][Spec Decode] Compact shared topk indices buffer after first MTP draft step (#47238)

2026-07-01 21:38:51 -07:00

multimodal

fix(security): prevent image decompression bomb OOM denial of service (#47010)

2026-06-30 09:39:22 +00:00

parser

[Bugfix][Frontend][gpt-oss] Return raw output when Harmony parser ends non-terminal (#47062)

2026-07-01 01:46:01 +00:00

platforms

[CPU][Perf]Added tanh AOR for faster gelu activations. (#44639)

2026-06-30 23:24:40 -07:00

plugins

Add vLLM library info to Hugging Face Hub requests (#43857)

2026-05-29 14:04:58 +00:00

profiler

[Refactor] Clean up log once scope="local" (#40540)

2026-04-22 16:42:43 -04:00

ray

[Ray] Propagate third-party env vars to Ray workers via prefix matching (#34383)

2026-02-17 01:08:42 -08:00

reasoning

[Model] Support Hy3 token suffix and JSON Schema array types (#47192)

2026-07-01 10:16:07 +00:00

renderers

fix(security): bound tokenizer work when explicit truncation_side is set (#47007)

2026-06-30 23:08:51 +08:00

third_party

[FlashMLA] Update FlashMLA to expose new arguments (#32810)

2026-01-21 22:02:39 -07:00

tokenizers

[Model] Support Unlimited OCR (#46564)

2026-06-27 23:09:18 -07:00

tool_parsers

[Bugfix][Tool Parser] poolside_v1: accept tool calls without newline after function name (#47311)

2026-07-02 12:08:37 -07:00

tracing

[Feature] OTEL tracing during loading (#31162)

2026-02-05 16:59:28 -08:00

transformers_utils

[Spec Decode] DSpark speculators checkpoint support (#47093)

2026-07-01 17:32:27 -07:00

triton_utils

[ROCm][Bugfix] Fix use_v2_model_runner inside Ray driver thread (#45998)

2026-06-24 08:41:56 +08:00

usage

replace cuda_device_count_stateless() to current_platform.device_count() (#37841)

2026-03-31 22:32:54 +08:00

utils

Xqa decode kernels (#43232)

2026-07-02 12:32:05 -07:00

v1

[MRV2] Draw 64-bit uniforms for fp32 Gumbel sampling

2026-07-03 06:22:59 +00:00

vllm_flash_attn

[Feat][1/N] CuTeDSL warmup infrastructure, FA4 MLA (#46182)

2026-06-30 12:17:34 -07:00

__init__.py

Remove bc-lint (#35274)

2026-02-26 03:01:01 -08:00

_aiter_ops.py

[ROCm][MLA] Fuse MLA q/kv RMSNorm + FP8 per-token quant in the FP8 attention path (#44977)

2026-07-02 13:00:41 +08:00

_custom_ops.py

Delete PagedAttention (#47361)

2026-07-02 12:31:26 -07:00

_xpu_ops.py

[BUGFIX][XPU] Update fa interface for compatibility (#45394)

2026-06-12 11:45:01 +00:00

collect_env.py

[XPU]Enhance environment collection for Intel XPU and optimize layout (#35698)

2026-04-13 12:51:46 +08:00

connections.py

[CI] Add retry with 4x backoff to HTTP fetches for transient failures (#37218)

2026-03-19 19:00:20 +00:00

env_override.py

[Perf] Add VLLM_TRITON_FORCE_FIRST_CONFIG to skip Triton autotuning (#42425)

2026-06-16 15:16:45 +02:00

envs.py

fix(security): prevent image decompression bomb OOM denial of service (#47010)

2026-06-30 09:39:22 +00:00

exceptions.py

[openapi server] log exception in exception handler(2/N) (#36201)

2026-03-10 20:16:30 -07:00

forward_context.py

[Optimization] Skip DP padding tokens in MoE (#46428)

2026-06-23 14:54:46 -07:00

logger.py

[Log] Log once in local node by default (#37568)

2026-03-19 12:04:59 -07:00

logits_process.py

…

logprobs.py

Make mypy opt-out instead of opt-in (#33205)

2026-01-29 09:12:26 +00:00

model_inspection.py

[UX] Add vLLM model inspection view (#29450)

2026-01-09 10:12:35 -07:00

outputs.py

Revert "[Core] Replace routing replay with device cache and async D2H pipeline" (#39917) (#42434)

2026-05-13 23:49:01 -07:00

pooling_params.py

[Bugfix] Reject matryoshka embedding dimensions above hidden size (#46313)

2026-06-22 10:16:35 +00:00

py.typed

…

sampling_params.py

[Bugfix][Structured Outputs] Reject degenerate structured_outputs that crash EngineCore (#45346)

2026-06-30 22:41:33 +00:00

scalar_type.py

Humming support for 2/3/5/6/7-bit pack-quantized weight-only inference (#46389)

2026-06-24 13:53:54 -04:00

scripts.py

…

sequence.py

[Core] Cleanup KVConnector handling with PP + fix MRV2 (#43732)

2026-05-28 13:12:03 -07:00

tasks.py

[Frontend] Remove frontend pooling multi task support. (#37861)

2026-04-21 12:27:44 +00:00

version.py

…