Commits vllm-project/vllm

[Bugfix] Add NaN masking to NVFP4 quantization to prevent output contamination

tlrmchlsmth committed 198 days ago

dcbedb76

[Bugfix] Revert NVFP4 check_tensor calls (incompatible with fullgraph)

tlrmchlsmth committed 198 days ago

Verified 0c534be7

[Bugfix] Make check_tensor fullgraph-safe and support FP8

tlrmchlsmth committed 198 days ago

Verified f408ad2b

[Bugfix] Fix check_tensor for torch.compile and FP8 dtypes

tlrmchlsmth committed 198 days ago

Verified 464b91bb

[Bugfix] Fix torch.compile graph break in NaN detector check_tensor

tlrmchlsmth committed 198 days ago

Verified d066cf30

[Kernel] Add NaN/Inf checks to NVFP4 linear pipeline

tlrmchlsmth committed 198 days ago

Verified ba42d161

[Kernel] Simplify NaN detector: remove post-crash KV scan

tlrmchlsmth committed 198 days ago

Verified 99e90c2a

[Kernel] Add pluggable NaN/Inf tensor checks to NaN detector

tlrmchlsmth committed 198 days ago

Verified 171eb482

[Kernel] Scan all KV cache blocks before crash on NaN detection

tlrmchlsmth committed 198 days ago

Verified 98502f1d

[Kernel] Crash on NaN/Inf detection in RMSNorm

tlrmchlsmth committed 198 days ago

Verified af85c7e9

[Kernel] Add KV cache block NaN checking to NaN detector

tlrmchlsmth committed 198 days ago

Verified 6d89d5a8

[Bugfix] Remove logger.warning_once() from RMSNorm hot path for torch.compile compatibility

tlrmchlsmth committed 198 days ago

a26a5cf8

Fix NaN from stale FP4 scale padding: torch.empty → torch.zeros

elvircrn committed 198 days ago

Verified d2afd40d

[Bugfix] Revert "Zero-init MLA attention output buffers to prevent NaN from CUDA graph padding (#37442)"

elvircrn committed 198 days ago

Verified eadf848e

[Test] Trim test_fused_quant_layernorm parametrization

tlrmchlsmth committed 198 days ago

Verified a827e390

[Test] Trim test_layernorm parametrization for faster CI

tlrmchlsmth committed 198 days ago

Verified f8499127

[Kernel] Add zero-cost NaN/Inf detection to RMSNorm kernels

tlrmchlsmth committed 198 days ago

Verified 0d98729a

Add NaN/Inf detection for NIXL KV cache transfers

tlrmchlsmth committed 198 days ago

Verified 7779fccd

[Bugfix] Zero-init MLA attention output buffers to prevent NaN from CUDA graph padding (#37442)

elvircrn committed 208 days ago

Verified ef2c4f77

[XPU]Unify xpu test dependencies in dockerfile.xpu (#36477)

1643661061leo committed 208 days ago

Verified 9dade5da

[Bugfix] Expand quantization method support in perf metrics (#37231)

thillai-c committed 208 days ago

Verified 828f862a

[Bugfix] Fix KV scales inconsistency in fp8 MLA & FlashInfer kv_cache_dtype "auto" leading to gibberish (#37054)

andylolu2 committed 208 days ago

Verified 577df69b

[Model Runner V2] Spec decode rejection sampler greedy support (#37238)

TheEpicDolphin committed 208 days ago

Verified 04244fd0

[Bugfix] Remove assertion for NVFP4 scale dynamic range (#37465)

mgoin committed 208 days ago

Verified 9482b0b0

[LoRA][BugFix] Fix skipped LoRA adapters for Mistral3 (#36928)

WoosukKwon committed 208 days ago

Verified 5bc1da14

fix(worker): optimize swap_states to copy only active token prefixes (#34733)

phiott committed 208 days ago

Verified 00910171

[V0 Deprecation] Deprecate virtual engine (#37195)

yewentao256 committed 208 days ago

Verified 0d81a1fe

chunk parakeet into 30s clips to prevent OOMs on long audios (#36671)

netanel-haber committed 208 days ago

Verified 6ae4c8d6

[Bugfix] Fix ROCm crash in qwen3_next multi-stream events (#36795) (#37427)

JartX committed 208 days ago

Verified a913b612

Fix models which use `layer_type_validation` for Transformers v5 (#37398)

hmellor committed 208 days ago

Verified 5ce2d10e