EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Eagle] Refactor eagle speculative decoding (#3986)

Browse Source 
Co-authored-by: Ke Bao <ISPObaoke@163.com>
This commit is contained in:
Ying Sheng
2025-03-05 08:06:07 -08:00
committed by GitHub
parent 5be8f1ed98
commit d3d4d76758
22 changed files with 670 additions and 352 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/bench_one_batch.py
View File

@@ -230,7 +230,7 @@ def extend(reqs, model_runner):
batch = ScheduleBatch.init_new(
reqs=reqs,
req_to_token_pool=model_runner.req_to_token_pool,
token_to_kv_pool=model_runner.token_to_kv_pool,
token_to_kv_pool_allocator=model_runner.token_to_kv_pool_allocator,
tree_cache=None,
model_config=model_runner.model_config,
enable_overlap=False,
@@ -326,7 +326,7 @@ def latency_test_run_once(
# Clear the pools.
model_runner.req_to_token_pool.clear()
model_runner.token_to_kv_pool.clear()
model_runner.token_to_kv_pool_allocator.clear()
measurement_results = {
"run_name": run_name,