EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Revert "[BugFix] Support setting tp=1 for the Eagle draft model to take effect (#5519)"(#5902)

Browse Source 
This reverts commit d886b81971. it breaks pd function

- vLLM version: v0.13.0
- vLLM main:
bde38c11df

Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
This commit is contained in:
zhaomingyu13
2026-01-14 20:55:10 +08:00
committed by GitHub
parent 2a6d95c389
commit 01805fbd7d
6 changed files with 11 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm_ascend/spec_decode/eagle_proposer.py
View File

@@ -130,27 +130,6 @@ class EagleProposer(VllmEagleProposer):
self.use_sparse = hasattr(vllm_config.model_config.hf_text_config,
"index_topk")
# NOTE:
# `draft_tensor_parallel_size` does not take effect for Eagle:
# the draft model uses the same TP size as the target model in practice.
# so we applied this patch to set tp=1 of draft model separately.
# Due to verification of `_verify_and_get_draft_tp` in vllm,
# the value of `draft_tensor_parallel_size` here will either be 1 separately
# or the same as target model.
# TODO(zhaomingyu13): If we want to adapt to the case where draft model tp
# is not 1 and differs from target model, this part should be rewritten.
if (vllm_config.parallel_config.tensor_parallel_size
!= self.speculative_config.draft_tensor_parallel_size):
tp_group = init_model_parallel_group(
[[get_world_group().rank]],
get_world_group().rank,
torch.distributed.get_backend(get_world_group().device_group),
use_message_queue_broadcaster=True,
group_name="tp",
)
self.tp_group_context = patch_tensor_parallel_group(tp_group)
else:
self.tp_group_context = nullcontext()
# TODO: Remove it when the bug of fx-graph is solved
self.maybe_eager_context: ContextManager[Any] = nullcontext()