Files

T

9c2d21046b [Bugfix][Spec Decode] Honour the draft's moe_backend on Model Runner V2 (#54788 )

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

2026-09-08 16:50:07 -07:00

metrics

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

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

[Perf] Avoid h2d copies from non-pinned CPU tensors (#54299 )

2026-08-29 12:06:30 +08:00

model_states

[1/2][Model Runner V2] DBO support, eager mode (#50945 )

2026-09-04 14:17:39 -07:00

pool

[Pooling] Improve BGE-M3 sync pooling throughput by up to 3.13% (#53464 )

2026-08-24 08:33:32 -04:00

sample

[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

spec_decode

[Bugfix][Spec Decode] Honour the draft's moe_backend on Model Runner V2 (#54788 )

2026-09-08 16:50:07 -07:00

__init__.py

…

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

[1/2][Model Runner V2] DBO support, eager mode (#50945 )

2026-09-04 14:17:39 -07:00

block_table.py

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

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

buffer_utils.py

[Core] Ensure memory is pinned prior to async h2d copy (#45424 )

2026-06-20 20:02:24 -07:00

cp_utils.py

[Spec decode] Support Kimi-K3 DCP with DSpark (#52188 )

2026-08-17 13:08:45 -07:00

cudagraph_utils.py

[Kimi Bug] Fix kimi k3 startup cuda graph issue with recoverSSM (#55774 )

2026-09-08 10:29:53 -04:00

dp_utils.py

[1/2][Model Runner V2] DBO support, eager mode (#50945 )

2026-09-04 14:17:39 -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

[Bugfix][PP] Never drop a decoding request from the sampled-token broadcast (#54436 )

2026-08-31 23:49:54 +00:00

kv_connector.py

[Core][KV Connector] Start async KV loads after the forward launch when no sync loads are scheduled (#53333 )

2026-08-28 21:58:00 +00:00

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

[Core][MRV2] Support eagle3 spec decode with pipeline parallel (#50514 )

2026-09-05 11:45:34 +08:00

pcp_manager.py

[Bugfix][PP] Never drop a decoding request from the sampled-token broadcast (#54436 )

2026-08-31 23:49:54 +00:00

pp_utils.py

[Core][MRV2] Support eagle3 spec decode with pipeline parallel (#50514 )

2026-09-05 11:45:34 +08:00

README.md

…

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

[1/2][Model Runner V2] DBO support, eager mode (#50945 )

2026-09-04 14:17:39 -07:00

warmup.py

[Bugfix] Defer adaptive verification until after kernel warmup (#55455 )

2026-09-06 11:31:40 -07:00

README.md

[Experimental] Model Runner V2

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