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

speculative: drop dead params/returns/no-ops (#24865)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-05-09 15:53:31 -07:00
committed by GitHub
parent b735ca178c
commit c95454b341
7 changed files with 15 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/cpu_graph_runner.py
+1 -1
View File
@@ -802,7 +802,7 @@ class CPUGraphRunner:
captured_forward_batch.encoder_lens[:raw_bs].copy_(
forward_batch.encoder_lens
)
if enable_num_token_non_padded(self.model_runner.server_args):
if enable_num_token_non_padded():
captured_forward_batch.num_token_non_padded.copy_(
forward_batch.num_token_non_padded
)
python/sglang/srt/model_executor/cuda_graph_runner.py
+2 -4
View File
@@ -987,7 +987,7 @@ class CudaGraphRunner:
# populate_from_forward_batch).
buffers.num_token_non_padded[...] = num_tokens
if (
enable_num_token_non_padded(self.model_runner.server_args)
enable_num_token_non_padded()
and self.require_gathered_buffer
and not self.nsa_enable_prefill_cp
):
@@ -1255,9 +1255,7 @@ class CudaGraphRunner:
require_gathered_buffer=self.require_gathered_buffer,
num_tokens_per_bs=self.num_tokens_per_bs,
nsa_enable_prefill_cp=self.nsa_enable_prefill_cp,
enable_num_token_non_padded_flag=enable_num_token_non_padded(
self.model_runner.server_args
),
enable_num_token_non_padded_flag=enable_num_token_non_padded(),
pp_proxy_tensors=pp_proxy_tensors,
)
python/sglang/srt/model_executor/forward_batch_info.py
+2 -2
View File
@@ -499,7 +499,7 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
)
num_tokens = len(batch.input_ids) if batch.input_ids is not None else 0
if enable_num_token_non_padded(model_runner.server_args):
if enable_num_token_non_padded():
ret.num_token_non_padded = torch.tensor(num_tokens, dtype=torch.int32).to(
device, non_blocking=True
)
@@ -1086,7 +1086,7 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
return self.tbo_split_seq_index is not None
def enable_num_token_non_padded(server_args):
def enable_num_token_non_padded():
return get_moe_expert_parallel_world_size() > 1
python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py
-1
View File
@@ -365,7 +365,6 @@ class EAGLEDraftExtendCudaGraphRunner:
num_accepted_drafts=num_accepted_drafts,
num_accepted_tokens=num_accepted_tokens,
)
spec_info.positions = None
self.deepep_adapter.capture(is_extend_in_batch=True)
python/sglang/srt/speculative/eagle_worker.py
+4 -6
View File
@@ -485,8 +485,8 @@ class EAGLEWorker(TpModelWorker):
set_time_batch(batch.reqs, "set_spec_draft_end_time", trace_only=True)
set_time_batch(batch.reqs, "set_spec_verify_start_time", trace_only=True)
logits_output, verify_output, model_worker_batch, can_run_cuda_graph = (
self.verify(batch, spec_info)
logits_output, verify_output, can_run_cuda_graph = self.verify(
batch, spec_info
)
if get_global_tracing_enabled():
@@ -527,9 +527,7 @@ class EAGLEWorker(TpModelWorker):
can_run_cuda_graph=can_run_cuda_graph,
)
def check_forward_draft_extend_after_decode(
self, batch: ScheduleBatch, verify_output: EagleVerifyOutput
):
def check_forward_draft_extend_after_decode(self, verify_output: EagleVerifyOutput):
local_need_forward = verify_output.unfinished_accept_tokens.shape[0] > 0
if not self.server_args.enable_dp_attention:
return local_need_forward
@@ -981,7 +979,7 @@ class EAGLEWorker(TpModelWorker):
)
batch.spec_info = res.next_draft_input
return logits_output, res, model_worker_batch, can_run_cuda_graph
return logits_output, res, can_run_cuda_graph
def _mamba_verify_update(
self,
python/sglang/srt/speculative/frozen_kv_mtp_worker.py
+2 -4
View File
@@ -447,9 +447,7 @@ class FrozenKVMTPWorker(TpModelWorker):
set_time_batch(batch.reqs, "set_spec_draft_end_time", trace_only=True)
set_time_batch(batch.reqs, "set_spec_verify_start_time", trace_only=True)
logits_output, verify_output, _, can_run_cuda_graph = self.verify(
batch, spec_info
)
logits_output, verify_output, can_run_cuda_graph = self.verify(batch, spec_info)
if get_global_tracing_enabled():
for idx, req in enumerate(batch.reqs):
@@ -771,4 +769,4 @@ class FrozenKVMTPWorker(TpModelWorker):
batch.spec_info = res.next_draft_input
del seq_lens_pre_verify
return logits_output, res, model_worker_batch, can_run_cuda_graph
return logits_output, res, can_run_cuda_graph
python/sglang/srt/speculative/multi_layer_eagle_worker.py
+4 -6
View File
@@ -272,8 +272,8 @@ class MultiLayerEagleWorker(TpModelWorker):
self.mtp_model_runner(0).tp_group
), speculative_moe_backend_context():
spec_info = self.draft(batch)
logits_output, verify_output, model_worker_batch, can_run_cuda_graph = (
self.verify(batch, spec_info)
logits_output, verify_output, can_run_cuda_graph = self.verify(
batch, spec_info
)
with self.draft_tp_context(
@@ -295,9 +295,7 @@ class MultiLayerEagleWorker(TpModelWorker):
can_run_cuda_graph=can_run_cuda_graph,
)
def check_forward_draft_extend_after_decode(
self, batch: ScheduleBatch, verify_output: EagleVerifyOutput
):
def check_forward_draft_extend_after_decode(self, verify_output: EagleVerifyOutput):
local_need_forward = verify_output.unfinished_accept_tokens.shape[0] > 0
if not self.server_args.enable_dp_attention:
return local_need_forward
@@ -593,7 +591,7 @@ class MultiLayerEagleWorker(TpModelWorker):
)
batch.spec_info = res.next_draft_input
return logits_output, res, model_worker_batch, can_run_cuda_graph
return logits_output, res, can_run_cuda_graph
def forward_draft_extend(
self,