vllm-project/vllm

Pull Requests Commits

Use -1 expert ID for padding and non-local tokens in DeepEP v2

tlrmchlsmth committed 120 days ago

Verified ae8c4457

Fix local→global expert ID conversion in DeepEP v2 decode path

tlrmchlsmth committed 121 days ago

Verified 45a0b6d7

Fix FP8 padding: use torch.where instead of masked_fill_

tlrmchlsmth committed 122 days ago

Verified 0f94cbef

Split DeepEP v2 into prefill/decode modes for memory vs cudagraph

tlrmchlsmth committed 122 days ago

Verified ed29278a

Update DeepEP v2 docstring with cudagraph design rationale

tlrmchlsmth committed 122 days ago

Verified 56a8767d

Refactor DeepEP v2 for cudagraph compatibility and simplified DBO

tlrmchlsmth committed 122 days ago

Verified 5a766bf3

Check runtime NCCL version instead of PyTorch compile-time constant

tlrmchlsmth committed 123 days ago

Verified 75149ae1

Fix NCCL version check: 2.30.4, not 4.30.4

tlrmchlsmth committed 123 days ago

Verified 67d26f78

Add DeepEP v2 lifecycle test to test_mnnvl_alltoall

tlrmchlsmth committed 123 days ago

Verified 019a41f0

Fix FP8 dispatch test to match production behavior

tlrmchlsmth committed 123 days ago

Verified 515e36da

Add DeepEP v2 (ElasticBuffer) all2all backend for MoE EP

tlrmchlsmth committed 123 days ago

Verified a2a4b00f

[CI] Add key field to all test_areas pipeline steps (#41201)

khluu committed 123 days ago

Verified 0ab67c02

[ROCm][Bugfix][GPTOSS]: fix input_ids and expert_map args for quark w4a8 gptoss (#41165)

Rohan138 committed 123 days ago

Verified 3795d7ac

[Ci][BugFix] Fix slow DP tests due to bad teardown logic (#41166)

njhill committed 123 days ago

Verified 18599bfd

[DSv4] Use `cvt` PTX for FP32->FP4 conversion (#41015)

gau-nernst committed 123 days ago

Verified 296741d0

[Bugfix][MLA] Size arange_buffer to max_num_batched_tokens to prevent CUDA IMA (#39277)

UranusSeven committed 123 days ago

Verified a966aaed

[ROCm] Add env flags to disable dynamic MXFP4 quant and enable AITER tuned GEMMs for Attention Projection Layers (#39987)

heachary committed 123 days ago

Verified 6841f5dc

[Bugfix][Compile] Fix gc.collect/empty_cache patch arity in CUDAGraphWrapper (#41235)

roikoren755 committed 123 days ago

Verified c2fb0133

Create tests/distributed/test_mnnvl_alltoall.py (#35241)

puririshi98 committed 123 days ago

Verified ccfb620c

[BUG] Two phase pause to prevent deadlock (#39366)

hao-aaron committed 123 days ago

Verified 0335316a

[ROCm][Bugfix]: W4A4 MOE using emulation instead of AITER on MXFP4-supported hardware (#41175)

Rohan138 committed 123 days ago

Verified 944e138b

[Perf][Spec Decode] Avoid per-step numpy allocation in prepare_next_t… (#41043)

wangluochao902 committed 123 days ago

Verified b58669cb

[Multimodal][Render] Skip mm processor initialization and warmup for text-only mode (#41246)

Isotr0py committed 123 days ago

Verified 1628239e

[CI] Add temperature to bfcl eval, default greedy (#41059)

yzong-rh committed 123 days ago

Verified 93da1fe9

[ROCm] Use quant_dtype in per_token_quant instead of hardcoded FP8 (#39121)

Bortlesboat committed 123 days ago

Verified 169988a3

[Feature]: IndexCache support for DSA models (#37735)

chaunceyjiang committed 123 days ago

Verified faab1895

Replace shape_invariants with simpler apprach in dynamic_arg_dims utilizing shape_id property. (#36194)

laithsakka committed 123 days ago

Verified 6f20f81c

Fix timeout when using LoRA adapters with Nemotron Super (#40916)

danisereb committed 123 days ago

Verified d1a75e30

[CI/Build] Enable FP8 on NVIDIA Thor (#39712)

DarkLight1337 committed 123 days ago

Verified 4a42aba3

better logging for large uncachable items (#41145)

h-avsha committed 123 days ago

Verified a80d6f15

Older