minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[NPU] bugs fix for Deepseek models (#19544)

Browse Source
This commit is contained in:
Michelle Wu
2026-02-28 17:26:15 +08:00
committed by GitHub
parent 1eb40d8d45
commit b7f13a7b73
2 changed files with 12 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/attention/mla_preprocess.py
+3 -2
View File
@@ -459,8 +459,9 @@ class NPUFusedMLAPreprocess(torch.nn.Module):
# assert self.quant_config and self.quant_config.get_name() == "modelslim" # assert self.quant_config and self.quant_config.get_name() == "modelslim"
# route by `qkv_a_proj` quant type as MTP layers can be unquantized # route by `qkv_a_proj` quant type as MTP layers can be unquantized
_is_w8a8 = ( _is_w8a8 = (
hasattr(self.qkv_a_proj.quant_method, "quant_config") hasattr(self.qkv_a_proj.quant_method, "quantization_config")
and self.qkv_a_proj.quant_method.quant_config.get_name() == "modelslim" and self.qkv_a_proj.quant_method.quantization_config.get_name()
== "modelslim"
) )
# with the mlaprolog enabled, the kv_b_proj layers are unquantized # with the mlaprolog enabled, the kv_b_proj layers are unquantized
_is_mlaprolog = hasattr(self.quant_config, "ignore") and any( _is_mlaprolog = hasattr(self.quant_config, "ignore") and any(
python/sglang/srt/managers/schedule_policy.py
+9 -9
View File
@@ -725,6 +725,15 @@ class PrefillAdder:
def add_one_req( def add_one_req(
self, req: Req, has_chunked_req: bool, truncation_align_size: Optional[int] self, req: Req, has_chunked_req: bool, truncation_align_size: Optional[int]
): ):
if (self.prefill_delayer_single_pass is not None) and (
not self.prefill_delayer_single_pass.negotiate_should_allow_prefill(
local_prefillable=True,
running_batch=self.running_batch.batch_size(),
max_prefill_bs=self.max_prefill_bs,
max_running_requests=self.max_running_requests,
)
):
return AddReqResult.OTHER
# TODO support cp with multiple requests # TODO support cp with multiple requests
# Enabling context parallelism currently presents precision issues; # Enabling context parallelism currently presents precision issues;
# therefore, the prefill-batch setting is temporarily set to 1. # therefore, the prefill-batch setting is temporarily set to 1.
@@ -771,15 +780,6 @@ class PrefillAdder:
if input_tokens >= self.rem_input_tokens and len(self.can_run_list) != 0: if input_tokens >= self.rem_input_tokens and len(self.can_run_list) != 0:
return AddReqResult.OTHER return AddReqResult.OTHER
if (self.prefill_delayer_single_pass is not None) and (
not self.prefill_delayer_single_pass.negotiate_should_allow_prefill(
local_prefillable=True,
running_batch=self.running_batch.batch_size(),
max_prefill_bs=self.max_prefill_bs,
max_running_requests=self.max_running_requests,
)
):
return AddReqResult.OTHER
if self.dllm_config is not None: if self.dllm_config is not None:
if self.rem_dllm_tokens <= 0: if self.rem_dllm_tokens <= 0: