Open-source activity

Wei-Cheng (Wayne) Chiu ships upstream.

GPU kernels, LLM serving, and correctness fixes across the open-source inference stack.

47 prs 22 completed 25 in-flight

Showing 47 of 47 pull requests

[Spec Decode] Fix adaptive verification startup on FA2 GPUs

vllm-project/vllm 94.1K

#621542 days ago

fix: give concurrent streams disjoint row_states buffers in multi-CTA top-k

flashinfer-ai/flashinfer 6.7K

#36253 days ago

test: bound peak memory in mxfp4 groupwise group GEMM test (32GB OOM)

flashinfer-ai/flashinfer 6.7K

#35942 weeks ago

fix: Weight autoscaling queue time by request count

triton-inference-server/server 11.1K

#8980last month

[Bugfix] Return HTTP 500 for non-streaming generate errors

vllm-project/vllm 94.1K

#491952 months ago

[Bugfix] Skip Qwen3 deepstack buffers without vision

vllm-project/vllm 94.1K

#493973 months ago

fix(planner): resolve aggregate workers by component type

ai-dynamo/dynamo 8.3K

#115783 months ago

Register wayne.is-a.dev

is-a-dev/register 11.6K

#441173 months ago

[Perf] kvcache: Fuse ordinary FP8 KV stores with shared conversion

sgl-project/sglang 37.4K

#316523 months ago

[#16196][fix] Detect MTP routed-expert layout separately

NVIDIA/TensorRT-LLM 14.9K

#162763 months ago

fix(attention): correct SM120 NVFP4 qk_correction layout, row-sum reduction, and lse

flashinfer-ai/flashinfer 6.7K

#38383 months ago

test(moe): align EP-offset tests and docs with global expert ids (gh #3547)

flashinfer-ai/flashinfer 6.7K

#35913 months ago

fix: make the cutlass MoE gemm profiler MXFP8-aware (autotune crash on MXFP8xMXFP8)

flashinfer-ai/flashinfer 6.7K

#36143 months ago

[Bugfix][Quantization] Upcast E8M0 scales for CUTLASS block FP8 and skip unaligned N on SM12x

vllm-project/vllm 94.1K

#479883 months ago

fix(moe): return bf16 expert_weights from trtllm_fp4_block_scale_moe (#3595)

flashinfer-ai/flashinfer 6.7K

#36443 months ago

[None][fix] Gate MoE BF16 to FP8 conversion on compiler support

NVIDIA/TensorRT-LLM 14.9K

#159344 months ago

fix(b12x): bucket MoE workspace allocation to stop kernel-cache thrash

flashinfer-ai/flashinfer 6.7K

#38364 months ago

feat(vllm): support NIXL push and pull connector names in PD

ai-dynamo/dynamo 8.3K

#112194 months ago

fix(common): make multimodal media loaders importable without torch

ai-dynamo/dynamo 8.3K

#111794 months ago

[Fix] model_config: Derive scaling in the KimiVL, MiniCPM3, and DeepseekVL2 MLA branches

sgl-project/sglang 37.4K

#298964 months ago

test: fix CUDA OOM in batch-prefill custom-mask test on 24GB CI GPUs

flashinfer-ai/flashinfer 6.7K

#36094 months ago

[Bugfix] Fix InternVL/InternVL3 LoRA loading TypeError in adapter fallback

InternLM/lmdeploy 8.1K

#46844 months ago

[Bugfix] Fix ImportError in get_chat_template for builtin chat-template names

InternLM/lmdeploy 8.1K

#46904 months ago

fix(mamba): reject SM120/SM121 in SSDCombined with a clear error

flashinfer-ai/flashinfer 6.7K

#36684 months ago

[SM120] Emit grouped NVFP4 (ue4m3) block-scaled GEMM in cutlass_library

NVIDIA/cutlass 10.6K

#33464 months ago

fix(trace): handle combined paged KV cache tensor in _get_tensor (#3685)

flashinfer-ai/flashinfer 6.7K

#36904 months ago

[Fix] [NVIDIA] trtllm_mla: Reject prefill and skip the decode counter on SM12x

sgl-project/sglang 37.4K

#288374 months ago

[Bugfix][Frontend] Use api_error for Anthropic streaming error events

vllm-project/vllm 94.1K

#461624 months ago

fix(logging): allow exporting OTLP traces without logs

ai-dynamo/dynamo 8.3K

#108064 months ago

[CuTeDSL] Fix make_ptr ZeroDivisionError on sub-byte dtypes

NVIDIA/cutlass 10.6K

#33344 months ago

[CuTeDSL] Fix cute.is_major crash on List[int] mode

NVIDIA/cutlass 10.6K

#33334 months ago

fix(kv-router): cancel in-flight recovery when worker is removed

ai-dynamo/dynamo 8.3K

#106164 months ago

[Bugfix] Fix double-counted max_q_seqlen in decode delta kv_seqlens

InternLM/lmdeploy 8.1K

#46854 months ago

fix(topk): eliminate multi-CTA radix top-k stream hangs on SM120/SM121

flashinfer-ai/flashinfer 6.7K

#36154 months ago

[Offload] Skip device-map entries with no local module in dispatch_with_map

vllm-project/compressed-tensors 330

#7374 months ago

fix(jit): quote -ccbin host compiler path so it works with spaces

flashinfer-ai/flashinfer 6.7K

#36374 months ago

fix(jit): cap ninja parallelism by available memory to avoid OOM (#3634)

flashinfer-ai/flashinfer 6.7K

#36364 months ago

[Fix] disaggregation: Pair ordinary MHA KV by layer ID across uneven prefill PP stages

sgl-project/sglang 37.4K

#282294 months ago

[PT2E][X86] Fall back to plain linear annotation for reused nn.Linear with fusable post op

#44804 months ago

[Bugfix] Reject structured outputs for diffusion decoders with a clear error

vllm-project/vllm 94.1K

#454684 months ago

[PT2E][X86] Extend reused-module fallback to nn.Conv2d (follow-up to #4480)

#44824 months ago

[Bugfix] Return the tokenizer from maybe_make_thread_pool so it survives pickling

vllm-project/vllm 94.1K

#454604 months ago

[Bugfix] Set type/role explicitly in streaming message_start event

vllm-project/vllm 94.1K

#453764 months ago

[Fix] hf_transformers: Apply rope overrides to text_config for VLM-format models

sgl-project/sglang 37.4K

#279834 months ago

[Bugfix] Skip Ray actor kill after the driver has disconnected

vllm-project/vllm 94.1K

#453284 months ago

[Bugfix] Reject NVFP4 MoE checkpoints with missing per-expert scales

vllm-project/vllm 94.1K

#453204 months ago

feat(operator): add RTX PRO 6000 Blackwell GPU SKU

ai-dynamo/dynamo 8.3K

#105794 months ago