Commit Graph

8050 Commits

Author SHA1 Message Date
Yaochen Han
577dbc4ab9 [4/N] Quantization Refactor: AWQ schemes and Kernel call and weight init split (#21126) 2026-04-30 14:51:01 +03:00
Qiaolin Yu
583929c0a1 fix the compatibility between --moe-dense-tp-size 1 and piecewise cuda graph (#23972) 2026-04-30 02:12:13 -07:00
Opher Lieber
99c0b62f1e allow requests with exactly context_len total tokens (#22546)
Co-authored-by: Ethan (Yusheng) Su <yushengsu.thu@gmail.com>
2026-04-30 01:12:06 -07:00
Ethan (Yusheng) Su
125f75db72 fix(lora): avoid CUDA graph-breaking scalar assignment in seg_indptr (#23738) 2026-04-30 01:11:45 -07:00
billishyahao
692979a8d9 [AMD] Support sdma path for moriep (#23929) 2026-04-29 23:57:00 -07:00
Shaojun Zhou
4f0b44c5c6 [fix] moss-vl: use Conv3dLayer and remove no-op flat_encoder_result (#23932) 2026-04-30 14:19:45 +08:00
kkyyxhll
936c9c2355 fix(qwen3_5): broadcast per-tensor scale in _make_packed_weight_loader for FP8 models (#23062) 2026-04-30 14:16:57 +08:00
Jay Thakur
bcb34da9f9 Add deterministic mode for XPU operations (#16793)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
2026-04-30 13:39:06 +08:00
Opher Lieber
c8c1c9261d LoRA support for qwen3.5 and nemotron3 (#23594)
Co-authored-by: Yanbin Jiang <jybsuper@gmail.com>
2026-04-29 21:51:53 -07:00
Mick
0b1fbdba15 [diffusion] CI: change ground truth upload path and improve publish script (#24120) 2026-04-30 12:26:10 +08:00
Yuxuan Zhang
d040333c95 [Bug Fix] missing index/KV transfer for MTP layer in NSA disaggregation (#23539) 2026-04-30 11:55:45 +08:00
yudian0504
2d2be5d7b2 [PD][Bugfix] fix mamba cache capping (#22462)
Co-authored-by: hzh0425 <hzh0425@apache.org>
Co-authored-by: yizhang2077 <1109276519@qq.com>
2026-04-30 10:57:55 +08:00
MingxuZh
62136073f9 pin the version of xgrammar to v0.1.32 (#24010)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
2026-04-30 10:13:08 +08:00
heziiop
3553fd0322 [NPU] add split_qkv_tp_rmsnorm_rope ops for minimax2 & fix eagle3 hidden states capture in dp attn mode (#23190) 2026-04-30 08:51:22 +08:00
Lianmin Zheng
e60c60eff0 [SWA] Fix missing mamba_indices parameter in cpu copy interface (#24026) 2026-04-29 17:33:38 -07:00
Kangyan-Zhou
6575aea128 [CI] Fix black formatting on main (unblocks PR #21247 lint) (#24093)
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-29 14:59:17 -07:00
Jimmy Shong
3d31ac2672 [Fix] FP8 Qwen3-Next quant error by removing fallback fused shards (#23973) 2026-04-29 17:33:47 -04:00
jsheng_Linkedin
850021378a [Score API] Hoist query placeholder scan and specialize PositionalEmbeds stacking (#23513)
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-29 13:51:53 -07:00
Qiaolin Yu
79dbfe4505 Use spec v2 by default (#21062) 2026-04-29 13:40:42 -07:00
Zhongdongming Dai
7389743d85 feat: Support modelexpress p2p RDMA transfer (#23105) 2026-04-29 12:57:40 -07:00
jsheng_Linkedin
db84a8ebbb [Model] Qwen3ForPooledOutput: forward get_input_embeddings to inner model (#23434) 2026-04-29 12:25:06 -07:00
Chang Min Bark
3272af2f00 [Apple Silicon] [MLX] MLX decode partial overlap scheduling for generation (async eval) (#22416)
Co-authored-by: R0CKSTAR <yeahdongcn@gmail.com>
Co-authored-by: Alex Nails <alex.nails@radixark.ai>
2026-04-29 12:21:14 -07:00
Xinyuan Tong
1376761841 fix(moe): repair dead import in fused_moe_native after MoE refactor (#24069) 2026-04-29 11:14:52 -07:00
Jeongho Shin
530b497a48 [BugFix] correct host leaf status check from evicted to backuped (#23537) 2026-04-29 10:26:00 -07:00
Xinyuan Tong
4cf109bbd1 debug followup (#24058) 2026-04-29 23:03:27 +08:00
Qi Yuhang
3f7c95d6cc [JIT Kernel][1/2]Migrate MXFP8 Group GEMM & Quant into JIT (#23833)
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
2026-04-29 22:50:09 +08:00
Mick
eeb7b5c433 [diffusion] fix: align encoder of flux klein with official (#24008) 2026-04-29 22:42:16 +08:00
Mick
e5c2a9b6cf [diffusion] fix: improve LTX2.3 reference accuracy controls (#24022) 2026-04-29 21:39:27 +08:00
Xinyuan Tong
1279ae0787 Bugfix (#24027) 2026-04-29 21:13:51 +08:00
AndyLi429
4c1eefca4f [NPU] ascend backend support qwen3 moe attention cp (#21685) 2026-04-29 19:25:17 +08:00
Zheng Wengang
ae0c036c24 [BugFix][EPD] fix embedding req_id transfer error (#23481) 2026-04-29 18:56:03 +08:00
jacky.cheng
180bb2624f [AMD] Fix CI RuntimeError: opentelemetry package is not installed (#23940)
Co-authored-by: Bingxu Chen <bingxche@amd.com>
2026-04-29 18:02:44 +08:00
hhwxw
d9270b8c6a fix(moe): relocate orphan tuned configs after #23019 (#24004) 2026-04-29 02:00:13 -07:00
Sam (Kesen Li)
73e93bebd6 [1/4] NVFP4 KV cache: quantization strategy abstraction and kernel (#21954) 2026-04-29 01:45:48 -07:00
danielafrimi
8327270c72 [Kernel] Support FlashInfer TRTLLM-Gen fused MoE for non-gated FP4 & FP8 (Nemotron) (#21321) 2026-04-29 01:28:21 -07:00
Liangsheng Yin
bd448e51bd [Spec] Split accept_length into num_accepted_drafts and num_accepted_tokens (#23962) 2026-04-29 00:02:22 -07:00
shuwenn
2c41ef4c93 [HiCache] feat: add draft KV cache backing for L2/L3 (#21125) 2026-04-28 23:47:31 -07:00
Mick
71d2227a78 [diffusion] CI: update ground truth with official output (#23714) 2026-04-29 13:51:31 +08:00
JieTang66
5a94088178 Fix typo: page_first_kv_spilt -> page_first_kv_split (#23983)
Co-authored-by: JieTang <tangjie66@huawei.com>
2026-04-29 12:08:05 +08:00
ZeyuanChen2000
23cfa9e4c8 [NPU] fix rope_theta get error for baichuan2-13b-chat model (#21543) 2026-04-29 12:01:03 +08:00
Yuhao Yang
b437f6be48 model: Nemotron-omni-v3-alias (#23857) 2026-04-29 11:08:23 +08:00
Qiaolin Yu
f57ec8d6ef [spec decoding] add extra attribute 'spec_hidden_size' (#23890) 2026-04-28 19:54:50 -07:00
Lianmin Zheng
2a771a40ac Add engine_type label to tokenizer manager metrics (#23978) 2026-04-28 19:52:58 -07:00
Rahul Vijayaraghavan
1aee04e7df [XPU] Support apply_router_weight_on_input for Llama4 for fused_experts (#22654)
merge this one as it is xpu only change.
2026-04-29 10:44:48 +08:00
ccullen-cert
1e14bd6f36 Fix for CVE-2026-5760 (#23660) 2026-04-28 19:39:44 -07:00
Hubert Lu
e5da200d0a [AMD] Fix Aiter RMSNorm layout handling (#23974) 2026-04-28 19:28:46 -07:00
Chandrakant Khandelwal
0ac23cffac Add intel_xpu as backend for GptOssForCausalLM, enabled for bf16 models with torch native backend (#12771)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
2026-04-29 10:21:52 +08:00
iridiumine
08699bb1b2 [NPU] Fix DeepEP LL dispatch BF16 flag and skip triton kernel on NPU for Qwen3.5 (#23815)
Co-authored-by: iridiumine <iridiumine@users.noreply.github.com>
2026-04-29 10:19:24 +08:00
Xiaoyu Zhang
13afe8acdf [codex] Enable Qwen3-Next MoE all-reduce fusion (#23619) 2026-04-29 09:11:35 +08:00
Lianmin Zheng
14b4e6fa69 Support --model as alias for --model-path in CLI (#23894) 2026-04-28 17:23:27 -07:00