https://prs.wayne.is-a.dev Discover Wei-Cheng (Wayne) Chiu's recent pull requests on GitHub Sun, 19 Jul 2026 13:23:47 GMT https://validator.w3.org/feed/docs/rss2.html https://github.com/jpmonette/feed en https://prs.wayne.is-a.dev/favicon.png https://prs.wayne.is-a.dev CC BY-NC-SA 4.0 2024 © Wei-Cheng (Wayne) Chiu https://github.com/is-a-dev/register/pull/44117 https://github.com/is-a-dev/register/pull/44117 Sat, 18 Jul 2026 22:59:28 GMT Register wayne.is-a.dev]]> https://github.com/sgl-project/sglang/pull/31652 https://github.com/sgl-project/sglang/pull/31652 Sat, 18 Jul 2026 05:07:15 GMT [JIT] Fuse FP8 KV-cache quantization into the store kernel]]> https://github.com/ai-dynamo/dynamo/pull/11578 https://github.com/ai-dynamo/dynamo/pull/11578 Sun, 12 Jul 2026 11:58:42 GMT fix(planner): resolve aggregate (mode=agg) DGDs via agg_worker_k8s_name]]> https://github.com/NVIDIA/TensorRT-LLM/pull/16276 https://github.com/NVIDIA/TensorRT-LLM/pull/16276 Sat, 11 Jul 2026 22:16:44 GMT [#16196][fix] Support MIXED_PRECISION experts-only NVFP4 DeepSeek-V4 checkpoints]]> https://github.com/vllm-project/vllm/pull/48199 https://github.com/vllm-project/vllm/pull/48199 Fri, 10 Jul 2026 02:16:56 GMT [Quantization] Fix misleading NVFP4 Marlin linear warning on FP4-native GPUs]]> https://github.com/vllm-project/vllm/pull/47988 https://github.com/vllm-project/vllm/pull/47988 Wed, 08 Jul 2026 11:45:05 GMT [Bugfix] Handle E8M0 block scales in CUTLASS and Triton FP8 linear kernels]]> https://github.com/flashinfer-ai/flashinfer/pull/3838 https://github.com/flashinfer-ai/flashinfer/pull/3838 Sat, 04 Jul 2026 15:16:07 GMT fix(attention): correct SM120 NVFP4 qk_correction layout, row-sum reduction, and lse]]> https://github.com/NVIDIA/TensorRT-LLM/pull/15934 https://github.com/NVIDIA/TensorRT-LLM/pull/15934 Sat, 04 Jul 2026 13:34:50 GMT [None][fix] Exclude SM12x from bf16x2 FP8 cvt specialization in moe_a2a]]> https://github.com/flashinfer-ai/flashinfer/pull/3836 https://github.com/flashinfer-ai/flashinfer/pull/3836 Sat, 04 Jul 2026 11:02:13 GMT fix(b12x): bucket MoE workspace allocation to stop kernel-cache thrash]]> https://github.com/flashinfer-ai/flashinfer/pull/3834 https://github.com/flashinfer-ai/flashinfer/pull/3834 Sat, 04 Jul 2026 05:16:13 GMT fix(sparse-mla-sm120): add TOPK=256 DSV4 prefill instantiation]]> https://github.com/ai-dynamo/dynamo/pull/11219 https://github.com/ai-dynamo/dynamo/pull/11219 Sat, 04 Jul 2026 04:48:55 GMT feat(vllm): support NixlPushConnector in PD coordination and NIXL detection]]> https://github.com/sgl-project/sglang/pull/30092 https://github.com/sgl-project/sglang/pull/30092 Sat, 04 Jul 2026 03:44:48 GMT fix(config): read num_nextn_predict_layers from the outer hf_config as fallback]]> https://github.com/flashinfer-ai/flashinfer/pull/3829 https://github.com/flashinfer-ai/flashinfer/pull/3829 Sat, 04 Jul 2026 02:44:25 GMT test: seed test_mm_fp4 inputs to fix flaky cos_sim threshold failures]]> https://github.com/vllm-project/vllm/pull/47577 https://github.com/vllm-project/vllm/pull/47577 Fri, 03 Jul 2026 22:12:33 GMT [MoE/b12x] Auto-select FLASHINFER_B12X NVFP4 MoE backend on SM120, reject EP deployments]]> https://github.com/flashinfer-ai/flashinfer/pull/3815 https://github.com/flashinfer-ai/flashinfer/pull/3815 Thu, 02 Jul 2026 23:24:06 GMT test(comm): order rank streams after input fills in dcp_alltoall tests]]> https://github.com/ai-dynamo/dynamo/pull/11179 https://github.com/ai-dynamo/dynamo/pull/11179 Thu, 02 Jul 2026 14:49:05 GMT fix(common): make multimodal media loaders importable without torch]]> https://github.com/sgl-project/sglang/pull/29902 https://github.com/sgl-project/sglang/pull/29902 Thu, 02 Jul 2026 05:52:06 GMT test(sgl-kernel): gate arch-specific tests that fail on SM120 (253 expected-unsupported failures)]]> https://github.com/sgl-project/sglang/pull/29896 https://github.com/sgl-project/sglang/pull/29896 Thu, 02 Jul 2026 05:07:30 GMT fix(config): derive scaling in the KimiVL, MiniCPM3, and DeepseekVL2 MLA branches]]> https://github.com/sgl-project/sglang/pull/29890 https://github.com/sgl-project/sglang/pull/29890 Thu, 02 Jul 2026 03:31:36 GMT fix(config): treat v_head_dim=0 as unset when deriving model shapes (deepseek-vl2-tiny warmup crash)]]> https://github.com/NVIDIA/cutlass/pull/3346 https://github.com/NVIDIA/cutlass/pull/3346 Tue, 23 Jun 2026 11:30:19 GMT [SM120] Emit grouped NVFP4 (ue4m3) block-scaled GEMM in cutlass_library]]> https://github.com/vllm-project/vllm/pull/46481 https://github.com/vllm-project/vllm/pull/46481 Tue, 23 Jun 2026 10:57:38 GMT [Kernel][SM120] NVFP4 grouped MoE: pingpong schedule at large per-exp…]]> https://github.com/flashinfer-ai/flashinfer/pull/3690 https://github.com/flashinfer-ai/flashinfer/pull/3690 Mon, 22 Jun 2026 21:56:35 GMT fix(trace): handle combined paged KV cache tensor in _get_tensor (#3685)]]> https://github.com/sgl-project/sglang/pull/28837 https://github.com/sgl-project/sglang/pull/28837 Sun, 21 Jun 2026 11:28:16 GMT fix(server_args): gate trtllm_mla prefill (not decode) off SM120/SM121]]> https://github.com/vllm-project/vllm/pull/46162 https://github.com/vllm-project/vllm/pull/46162 Fri, 19 Jun 2026 13:43:25 GMT [Bugfix] Forward upstream error message in Anthropic streaming converter (#46028)]]> https://github.com/ai-dynamo/dynamo/pull/10806 https://github.com/ai-dynamo/dynamo/pull/10806 Fri, 19 Jun 2026 12:03:25 GMT fix(logging): allow exporting OTLP traces without logs]]> https://github.com/NVIDIA/cutlass/pull/3334 https://github.com/NVIDIA/cutlass/pull/3334 Fri, 19 Jun 2026 08:32:00 GMT [CuTeDSL] Fix make_ptr ZeroDivisionError on sub-byte dtypes]]> https://github.com/NVIDIA/cutlass/pull/3333 https://github.com/NVIDIA/cutlass/pull/3333 Fri, 19 Jun 2026 08:11:07 GMT [CuTeDSL] Fix cute.is_major crash on List[int] mode]]> https://github.com/InternLM/lmdeploy/pull/4690 https://github.com/InternLM/lmdeploy/pull/4690 Fri, 19 Jun 2026 06:24:52 GMT [Bugfix] Fix ImportError in get_chat_template for builtin chat-template names]]> https://github.com/flashinfer-ai/flashinfer/pull/3670 https://github.com/flashinfer-ai/flashinfer/pull/3670 Thu, 18 Jun 2026 02:35:08 GMT fix(moe): don't mark trtllm-gen FP4 MoE backend supported on SM120/SM121]]> https://github.com/flashinfer-ai/flashinfer/pull/3669 https://github.com/flashinfer-ai/flashinfer/pull/3669 Thu, 18 Jun 2026 02:23:24 GMT fix(mla): don't recommend backend='cutlass' on SM120/SM121 MLA fallback]]> https://github.com/flashinfer-ai/flashinfer/pull/3668 https://github.com/flashinfer-ai/flashinfer/pull/3668 Thu, 18 Jun 2026 02:19:52 GMT fix(mamba): reject SM120/SM121 in SSDCombined with a clear error]]> https://github.com/InternLM/lmdeploy/pull/4685 https://github.com/InternLM/lmdeploy/pull/4685 Tue, 16 Jun 2026 22:14:54 GMT [Bugfix] Fix double-counted max_q_seqlen in decode delta kv_seqlens]]> https://github.com/InternLM/lmdeploy/pull/4684 https://github.com/InternLM/lmdeploy/pull/4684 Tue, 16 Jun 2026 13:54:38 GMT [Bugfix] Fix InternVL/InternVL3 LoRA loading TypeError in adapter fallback]]> https://github.com/LMCache/LMCache/pull/3722 https://github.com/LMCache/LMCache/pull/3722 Tue, 16 Jun 2026 13:38:57 GMT [Bugfix] Clear reqs_status on async lookup timeout to prevent KeyError on recall]]> https://github.com/vllm-project/compressed-tensors/pull/737 https://github.com/vllm-project/compressed-tensors/pull/737 Tue, 16 Jun 2026 13:01:19 GMT [Offload] Skip device-map entries with no local module in dispatch_with_map]]> https://github.com/vllm-project/vllm/pull/45811 https://github.com/vllm-project/vllm/pull/45811 Tue, 16 Jun 2026 11:52:09 GMT [Bugfix] Don't fail model load for multimodal models with no language model (#45806)]]> https://github.com/vllm-project/vllm/pull/45754 https://github.com/vllm-project/vllm/pull/45754 Tue, 16 Jun 2026 01:19:10 GMT [Bugfix] DiffusionGemma: only pop a request's logprobs when it commits (#45689)]]> https://github.com/flashinfer-ai/flashinfer/pull/3644 https://github.com/flashinfer-ai/flashinfer/pull/3644 Mon, 15 Jun 2026 19:19:28 GMT fix(moe): return bf16 expert_weights from trtllm_fp4_block_scale_moe (#3595)]]> https://github.com/sgl-project/sglang/pull/28257 https://github.com/sgl-project/sglang/pull/28257 Mon, 15 Jun 2026 07:12:53 GMT [Fix] Update speculative draft weights on update_weights_from_distributed]]> https://github.com/flashinfer-ai/flashinfer/pull/3637 https://github.com/flashinfer-ai/flashinfer/pull/3637 Mon, 15 Jun 2026 04:08:42 GMT fix(jit): quote -ccbin host compiler path so it works with spaces]]> https://github.com/flashinfer-ai/flashinfer/pull/3636 https://github.com/flashinfer-ai/flashinfer/pull/3636 Mon, 15 Jun 2026 03:16:59 GMT fix(jit): cap ninja parallelism by available memory to avoid OOM (#3634)]]> https://github.com/sgl-project/sglang/pull/28229 https://github.com/sgl-project/sglang/pull/28229 Mon, 15 Jun 2026 03:10:31 GMT fix: correct MHA PP KV-pointer slicing for uneven full-attn splits (#27740)]]> https://github.com/vllm-project/vllm/pull/45530 https://github.com/vllm-project/vllm/pull/45530 Sat, 13 Jun 2026 14:22:07 GMT [Bugfix] Bounds-check moe_permute reverse-map write (#45492)]]> https://github.com/sgl-project/sglang/pull/28137 https://github.com/sgl-project/sglang/pull/28137 Sat, 13 Jun 2026 08:23:59 GMT [Kernel] Add SM120/SM121 dispatch for int8_scaled_mm]]> https://github.com/pytorch/ao/pull/4482 https://github.com/pytorch/ao/pull/4482 Sat, 13 Jun 2026 07:36:21 GMT [PT2E][X86] Extend reused-module fallback to nn.Conv2d (follow-up to #4480)]]> https://github.com/sgl-project/sglang/pull/28125 https://github.com/sgl-project/sglang/pull/28125 Sat, 13 Jun 2026 06:02:15 GMT feat(moe): add SM120/SM121 dispatch for fp8_blockwise_scaled_grouped_mm]]> https://github.com/flashinfer-ai/flashinfer/pull/3625 https://github.com/flashinfer-ai/flashinfer/pull/3625 Fri, 12 Jun 2026 22:41:37 GMT fix: give concurrent streams disjoint row_states buffers in multi-CTA top-k]]> https://github.com/vllm-project/vllm/pull/45468 https://github.com/vllm-project/vllm/pull/45468 Fri, 12 Jun 2026 22:37:45 GMT [Bugfix] Reject structured outputs for diffusion decoders with a clear error]]> https://github.com/pytorch/ao/pull/4480 https://github.com/pytorch/ao/pull/4480 Fri, 12 Jun 2026 22:15:02 GMT [PT2E][X86] Fall back to plain linear annotation for reused nn.Linear with fusable post op]]> https://github.com/vllm-project/vllm/pull/45460 https://github.com/vllm-project/vllm/pull/45460 Fri, 12 Jun 2026 22:01:24 GMT [Bugfix] Return the tokenizer from maybe_make_thread_pool so it survives pickling]]> https://github.com/sgl-project/sglang/pull/28077 https://github.com/sgl-project/sglang/pull/28077 Fri, 12 Jun 2026 17:57:25 GMT Add tuned fused-MoE Triton config for Gemma-4 FP8 on RTX PRO 6000 Blackwell (SM120)]]> https://github.com/flashinfer-ai/flashinfer/pull/3615 https://github.com/flashinfer-ai/flashinfer/pull/3615 Fri, 12 Jun 2026 11:25:00 GMT fix(topk): eliminate multi-CTA radix top-k stream hangs on SM120/SM121]]> https://github.com/vllm-project/vllm/pull/45406 https://github.com/vllm-project/vllm/pull/45406 Fri, 12 Jun 2026 11:23:05 GMT [BugFix] Don't strand parked async-KV-load requests behind an unschedulable queue head]]> https://github.com/sgl-project/sglang/pull/28038 https://github.com/sgl-project/sglang/pull/28038 Fri, 12 Jun 2026 09:20:06 GMT Clamp default fused-MoE Triton configs to the device shared memory limit (SM120/SM121)]]> https://github.com/vllm-project/vllm/pull/45393 https://github.com/vllm-project/vllm/pull/45393 Fri, 12 Jun 2026 08:59:12 GMT [Kernel] Resurrect rotted gather-cache OOB test + defensive block-table bound-checks]]> https://github.com/flashinfer-ai/flashinfer/pull/3614 https://github.com/flashinfer-ai/flashinfer/pull/3614 Fri, 12 Jun 2026 08:32:04 GMT fix: make the cutlass MoE gemm profiler MXFP8-aware (autotune crash on MXFP8xMXFP8)]]> https://github.com/vllm-project/vllm/pull/45384 https://github.com/vllm-project/vllm/pull/45384 Fri, 12 Jun 2026 07:51:41 GMT [Bugfix] Fix int32 overflow in concat_mla_q flat warp index]]> https://github.com/flashinfer-ai/flashinfer/pull/3609 https://github.com/flashinfer-ai/flashinfer/pull/3609 Fri, 12 Jun 2026 07:46:21 GMT test: fix CUDA OOM in batch-prefill custom-mask test on 24GB CI GPUs]]> https://github.com/sgl-project/sglang/pull/28015 https://github.com/sgl-project/sglang/pull/28015 Fri, 12 Jun 2026 07:06:18 GMT Fix BF16 RL weight update with non-routed flashinfer_trtllm MoE backend]]> https://github.com/vllm-project/vllm/pull/45376 https://github.com/vllm-project/vllm/pull/45376 Fri, 12 Jun 2026 06:53:27 GMT [Bugfix] Set type/role explicitly in streaming message_start event]]> https://github.com/sgl-project/sglang/pull/27983 https://github.com/sgl-project/sglang/pull/27983 Fri, 12 Jun 2026 00:18:49 GMT fix: apply rope overrides to text_config for VLM-format models]]> https://github.com/vllm-project/vllm/pull/45334 https://github.com/vllm-project/vllm/pull/45334 Thu, 11 Jun 2026 23:46:49 GMT [Bugfix] Make block-fp8 fused MoE test tolerance noise-aware]]> https://github.com/vllm-project/vllm/pull/45328 https://github.com/vllm-project/vllm/pull/45328 Thu, 11 Jun 2026 22:39:25 GMT [Bugfix] Skip Ray actor kill after the driver has disconnected]]> https://github.com/flashinfer-ai/flashinfer/pull/3594 https://github.com/flashinfer-ai/flashinfer/pull/3594 Thu, 11 Jun 2026 22:26:14 GMT test: bound peak memory in mxfp4 groupwise group GEMM test (32GB OOM)]]> https://github.com/vllm-project/vllm/pull/45326 https://github.com/vllm-project/vllm/pull/45326 Thu, 11 Jun 2026 22:08:48 GMT [Bugfix] Abort requests that arrive while the engine is asleep]]> https://github.com/sgl-project/sglang/pull/27963 https://github.com/sgl-project/sglang/pull/27963 Thu, 11 Jun 2026 21:43:33 GMT test: cover diffkv (qk192/v128) + attention-sink triton attention kernels]]> https://github.com/sgl-project/sglang/pull/27962 https://github.com/sgl-project/sglang/pull/27962 Thu, 11 Jun 2026 21:42:54 GMT fix(attention): sm120 shared-memory-safe block size in context_attention_fwd]]> https://github.com/vllm-project/vllm/pull/45320 https://github.com/vllm-project/vllm/pull/45320 Thu, 11 Jun 2026 21:20:11 GMT [Bugfix] Reject NVFP4 MoE checkpoints with missing per-expert scales]]> https://github.com/flashinfer-ai/flashinfer/pull/3592 https://github.com/flashinfer-ai/flashinfer/pull/3592 Thu, 11 Jun 2026 21:10:52 GMT fix(moe): thread calibrated NVFP4 global scales through the unified MoE API]]> https://github.com/flashinfer-ai/flashinfer/pull/3591 https://github.com/flashinfer-ai/flashinfer/pull/3591 Thu, 11 Jun 2026 20:50:16 GMT test(moe): align EP-offset tests and docs with global expert ids (gh #3547)]]> https://github.com/ai-dynamo/dynamo/pull/10616 https://github.com/ai-dynamo/dynamo/pull/10616 Thu, 11 Jun 2026 20:40:55 GMT fix(kv-router): cancel in-flight recovery when worker is removed]]> https://github.com/sgl-project/sglang/pull/27916 https://github.com/sgl-project/sglang/pull/27916 Thu, 11 Jun 2026 10:56:51 GMT Fix MiMo-V2 expert routing on TopK-bypassing MoE backends]]> https://github.com/flashinfer-ai/flashinfer/pull/3577 https://github.com/flashinfer-ai/flashinfer/pull/3577 Thu, 11 Jun 2026 05:34:37 GMT test: cover XQA decode head_dim=256 with small page sizes (#2638)]]> https://github.com/ai-dynamo/dynamo/pull/10579 https://github.com/ai-dynamo/dynamo/pull/10579 Thu, 11 Jun 2026 05:09:55 GMT feat: add RTX PRO 6000 Blackwell GPU SKU]]> https://github.com/NVIDIA/TensorRT-LLM/pull/15155 https://github.com/NVIDIA/TensorRT-LLM/pull/15155 Tue, 09 Jun 2026 09:14:45 GMT [#14146][bugfix] Fix CuteDSL MoE ghost-token and global-index bugs]]> https://github.com/NVIDIA/TensorRT-LLM/pull/15154 https://github.com/NVIDIA/TensorRT-LLM/pull/15154 Tue, 09 Jun 2026 09:03:39 GMT [None][bugfix] Guard e_score_correction_bias load for DeepSeek-V2-Lite]]>