Commit 04d573dc authored by yangql's avatar yangql
Browse files

Merge branch 'v0.9.2-dev-ds_auto_21' into 'v0.9.2-dev-ds_auto_12.29'

[feat]解决deepep auto模式 mtp >1 时卡住

See merge request dcutoolkit/deeplearing/vllm!326
parents 059b1961 37d2b35d
...@@ -566,15 +566,15 @@ class EagleProposer: ...@@ -566,15 +566,15 @@ class EagleProposer:
) )
if self.dp_size > 1 and self.enable_expert_parallel and self.num_speculative_tokens > 1: if self.dp_size > 1 and self.enable_expert_parallel and self.num_speculative_tokens > 1:
num_tokens = 1
for _ in range(self.num_speculative_tokens - 1): for _ in range(self.num_speculative_tokens - 1):
with set_forward_context(attn_metadata, with set_forward_context(attn_metadata,
self.vllm_config, self.vllm_config,
num_tokens=num_input_tokens, num_tokens=num_tokens,):
num_tokens_across_dp=num_tokens_across_dp):
self.model( self.model(
self.input_ids[:num_input_tokens], self.input_ids[:num_tokens],
self.positions[:num_input_tokens], self.positions[:num_tokens],
self.hidden_states[:num_input_tokens], self.hidden_states[:num_tokens],
) )
def validate_same_kv_cache_group(self, def validate_same_kv_cache_group(self,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment