chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Core] Replace server_args mutation hack with explicit MemoryPoolConfig for draft worker init (#20183)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-03-09 11:45:54 -07:00
committed by GitHub
parent ecca8c553d
commit ffb4b6f4c1
9 changed files with 93 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
python/sglang/srt/speculative/eagle_worker.py
View File

@@ -152,6 +152,7 @@ class EAGLEWorker(TpModelWorker):
is_draft_worker=True,
req_to_token_pool=self.req_to_token_pool,
token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
memory_pool_config=target_worker.model_runner.memory_pool_config,
)
embed, head = self.target_worker.model_runner.model.get_embed_and_head()

1
python/sglang/srt/speculative/eagle_worker_v2.py
View File

@@ -145,6 +145,7 @@ class EagleDraftWorker(BaseDraftWorker):
is_draft_worker=True,
req_to_token_pool=self.req_to_token_pool,
token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
memory_pool_config=target_worker.model_runner.memory_pool_config,
)
# Alias for better readability

1
python/sglang/srt/speculative/multi_layer_eagle_worker.py
View File

@@ -142,6 +142,7 @@ class MultiLayerEagleWorker(TpModelWorker):
is_draft_worker=True,
req_to_token_pool=self.req_to_token_pool,
token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
memory_pool_config=target_worker.model_runner.memory_pool_config,
is_multi_layer_eagle=True,
)

1
python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py
View File

@@ -126,6 +126,7 @@ class MultiLayerEagleDraftWorker(BaseDraftWorker):
is_draft_worker=True,
req_to_token_pool=self.req_to_token_pool,
token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
memory_pool_config=target_worker.model_runner.memory_pool_config,
is_multi_layer_eagle=True,
)

1
python/sglang/srt/speculative/standalone_worker.py
View File

@@ -86,6 +86,7 @@ class StandaloneWorker(EAGLEWorker):
is_draft_worker=True,
req_to_token_pool=self.req_to_token_pool,
token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
memory_pool_config=target_worker.model_runner.memory_pool_config,
)
# Init attention backend and cuda graphs

1
python/sglang/srt/speculative/standalone_worker_v2.py
View File

@@ -99,6 +99,7 @@ class StandaloneDraftWorker(EagleDraftWorker):
is_draft_worker=True,
req_to_token_pool=self.req_to_token_pool,
token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
memory_pool_config=target_worker.model_runner.memory_pool_config,
)
# Alias for better readability