Files

T

zack 0136df94b0 [Bugfix][Spec Decode][MoE] Avoid uninitialized EPLB state in DeepSeek V4.1 DSpark drafter (#56387 )

Signed-off-by: zack <51604064+luoyuctl@users.noreply.github.com>

2026-09-15 14:35:47 +00:00

metrics

[Model Runner V2] Support num NaNs in logits (#30187 )

2025-12-09 10:00:49 -08:00

[Core][Model] Unify XD-RoPE into M-RoPE and derive the channel count (#56078 )

2026-09-09 09:23:30 -07:00

model_states

[PCP][DCP] Enable PCP+DCP on sparse-MLA models (#56157 )

2026-09-12 23:24:44 +00:00

pool

[Pooling] Support prompt embeddings in MRV2 decoder pooling (#56639 )

2026-09-13 05:49:45 +00:00

sample

[MRV2] Revert explicit Triton JIT warmup migration (#56654 )

2026-09-13 04:07:56 +00:00

spec_decode

[Bugfix][Spec Decode][MoE] Avoid uninitialized EPLB state in DeepSeek V4.1 DSpark drafter (#56387 )

2026-09-15 14:35:47 +00:00

__init__.py

GPU Model Runner V2 (#25266 )

2025-11-21 08:20:55 -08:00

async_utils.py

[Perf][Model Runner V2] Compact sampling masks on GPU instead of unpacking the full-vocab bitmask on CPU (#54901 )

2026-09-03 19:03:36 -07:00

attn_utils.py

[Warmup] Gemma 4 de-JITification (#55768 )

2026-09-13 16:47:21 -04:00

block_table.py

[Model] Support Qwen3.8-Flash-Next (#53896 )

2026-08-31 05:57:56 +00:00

buffer_utils.py

[Model Runner V2] Tolerate lack of pinned memory support (#56908 )

2026-09-14 21:28:32 -07:00

cp_utils.py

[Bugfix][MRV2] Initialize DCP metadata after batch partitioning (#55212 )

2026-09-10 06:50:09 +00:00

cudagraph_utils.py

[Feature][PCP] Support decode-only FULL CUDA graphs (#53867 )

2026-09-15 16:27:46 +08:00

dp_utils.py

[2/2][Model Runner V2] FULL CUDA graph capture for microbatched steps (DBO) (#51700 )

2026-09-13 22:42:33 -07:00

ec_connector.py

[EPD] Add ECMooncakeConnector for encoder cache over Mooncake TransferEngine (#41567 )

2026-09-07 06:21:43 +00:00

eplb_utils.py

[Elastic EP] Reduce eager-mode reconfiguration downtime (#51885 )

2026-08-19 02:13:06 +00:00

input_batch.py

[Feature][PCP] Support decode-only FULL CUDA graphs (#53867 )

2026-09-15 16:27:46 +08:00

kv_connector.py

[3/N] HiSparse: host-resident sparse-MLA decode hot-buffering (#53781 )

2026-09-12 00:36:47 -04:00

lora_utils.py

[BugFix][MRV2] Don't create dummy requests longer than max_model_len (#49751 )

2026-07-27 02:04:20 +00:00

model_runner.py

[Feature][PCP] Support decode-only FULL CUDA graphs (#53867 )

2026-09-15 16:27:46 +08:00

pcp_manager.py

[Feature][PCP] Support decode-only FULL CUDA graphs (#53867 )

2026-09-15 16:27:46 +08:00

pp_utils.py

[Bugfix][V2] record_stream idx_mapping in the PP draft broadcast (#55745 )

2026-09-09 12:55:56 -07:00

README.md

GPU Model Runner V2 (#25266 )

2025-11-21 08:20:55 -08:00

shutdown.py

[MRV2] Add shutdown() method (#41297 )

2026-05-02 21:06:30 -07:00

states.py

[Model Runner V2][Spec Decode] Add multi-layer MTP speculator (#48892 )

2026-07-30 15:54:15 -07:00

structured_outputs.py

[Bugfix][Spec Decode][Structured Output] DSpark: fix the grammar bitmask mapping when the draft budget is zero (#52436 )

2026-08-16 11:09:23 +00:00

ubatch_utils.py

[2/2][Model Runner V2] FULL CUDA graph capture for microbatched steps (DBO) (#51700 )

2026-09-13 22:42:33 -07:00

warmup.py

[3/N] HiSparse: host-resident sparse-MLA decode hot-buffering (#53781 )

2026-09-12 00:36:47 -04:00

README.md

[Experimental] Model Runner V2

This directory contains the new model runner which is under active development. Ping Woosuk Kwon for any changes.