[MoE Refactor] Remove MoE DP chunking (#39107)

Signed-off-by: Bill Nell <bnell@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
bnellnm
2026-04-14 09:48:05 -04:00
committed by GitHub
co-authored by Robert Shaw
parent f7e62e3d66
commit e1e318af01
20 changed files with 76 additions and 389 deletions
+16 -7
View File
@@ -1588,9 +1588,6 @@ class EngineArgs:
self._check_feature_supported()
self._set_default_chunked_prefill_and_prefix_caching_args(model_config)
self._set_default_max_num_seqs_and_batched_tokens_args(
usage_context, model_config
)
self._set_default_reasoning_config_args()
sliding_window: int | None = None
if not is_interleaved(model_config.hf_text_config):
@@ -1846,6 +1843,12 @@ class EngineArgs:
target_parallel_config=parallel_config,
)
self._set_default_max_num_seqs_and_batched_tokens_args(
usage_context,
model_config,
parallel_config,
)
assert self.max_num_batched_tokens is not None, (
"max_num_batched_tokens must be set by this point"
)
@@ -2244,6 +2247,7 @@ class EngineArgs:
self,
usage_context: UsageContext | None,
model_config: ModelConfig,
parallel_config: ParallelConfig,
):
world_size = self.pipeline_parallel_size * self.tensor_parallel_size
(
@@ -2255,10 +2259,15 @@ class EngineArgs:
orig_max_num_seqs = self.max_num_seqs
if self.max_num_batched_tokens is None:
self.max_num_batched_tokens = default_max_num_batched_tokens.get(
usage_context,
SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS,
)
if parallel_config.use_batched_dp_moe:
self.max_num_batched_tokens = (
SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS_FOR_BATCHED_DP
)
else:
self.max_num_batched_tokens = default_max_num_batched_tokens.get(
usage_context,
SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS,
)
if self.max_num_seqs is None:
self.max_num_seqs = default_max_num_seqs.get(