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

clean up gemlite usage (#14444)

Browse Source
This commit is contained in:
Minglei Zhu
2025-12-04 21:52:56 -08:00
committed by GitHub
parent 80a575e4e8
commit b76e303e6a
3 changed files with 0 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/model_executor/cuda_graph_runner.py
View File

@@ -51,7 +51,6 @@ from sglang.srt.layers.dp_attention import (
from sglang.srt.layers.logits_processor import LogitsProcessorOutput
from sglang.srt.layers.moe.token_dispatcher.deepep import DeepEPBuffer
from sglang.srt.layers.moe.utils import get_deepep_mode, get_moe_a2a_backend
from sglang.srt.layers.torchao_utils import save_gemlite_cache
from sglang.srt.model_executor.forward_batch_info import (
CaptureHiddenMode,
ForwardBatch,
@@ -489,9 +488,6 @@ class CudaGraphRunner:
self.graphs[key] = graph
self.output_buffers[key] = output_buffers
# Save gemlite cache after each capture
save_gemlite_cache()
# Trigger CUDA graph capture for specific shapes.
# Capture the large shapes first so that the smaller shapes
# can reuse the memory pool allocated for the large shapes.

4
python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
View File

@@ -45,7 +45,6 @@ from sglang.srt.layers.dp_attention import (
)
from sglang.srt.layers.logits_processor import LogitsProcessorOutput
from sglang.srt.layers.pooler import EmbeddingPoolerOutput
from sglang.srt.layers.torchao_utils import save_gemlite_cache
from sglang.srt.model_executor.forward_batch_info import (
CaptureHiddenMode,
ForwardBatch,
@@ -396,9 +395,6 @@ class PiecewiseCudaGraphRunner:
with set_compiled(True):
self.capture_one_batch_size(num_tokens)
# Save gemlite cache after each capture
save_gemlite_cache()
def capture_one_batch_size(self, num_tokens: int):
bs = 1