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

fix retrive -> retrieve typo (#23503)

Browse Source 
Co-authored-by: SoluMilken <19161836+solumilken@users.noreply.github.com>
This commit is contained in:
Liangsheng Yin
2026-04-22 16:35:04 -07:00
committed by GitHub
co-authored by SoluMilken
parent 917d2aa1dc
commit f611dd24f1
16 changed files with 177 additions and 163 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/batch_overlap/two_batch_overlap.py
+19 -17
View File
@@ -220,24 +220,26 @@ def split_spec_info(
positions = spec_info.positions[start_token_index:end_token_index] positions = spec_info.positions[start_token_index:end_token_index]
else: else:
positions = None positions = None
if spec_info.retrive_index is not None: if spec_info.retrieve_index is not None:
retrive_index = spec_info.retrive_index[start_seq_index:end_seq_index] retrieve_index = spec_info.retrieve_index[start_seq_index:end_seq_index]
else: else:
retrive_index = None retrieve_index = None
if spec_info.retrive_next_token is not None: if spec_info.retrieve_next_token is not None:
retrive_next_token = spec_info.retrive_next_token[start_seq_index:end_seq_index] retrieve_next_token = spec_info.retrieve_next_token[
else:
retrive_next_token = None
if spec_info.retrive_next_sibling is not None:
retrive_next_sibling = spec_info.retrive_next_sibling[
start_seq_index:end_seq_index start_seq_index:end_seq_index
] ]
else: else:
retrive_next_sibling = None retrieve_next_token = None
if spec_info.retrive_cum_len is not None: if spec_info.retrieve_next_sibling is not None:
retrive_cum_len = spec_info.retrive_cum_len[start_seq_index:end_seq_index] retrieve_next_sibling = spec_info.retrieve_next_sibling[
start_seq_index:end_seq_index
]
else: else:
retrive_cum_len = None retrieve_next_sibling = None
if spec_info.retrieve_cum_len is not None:
retrieve_cum_len = spec_info.retrieve_cum_len[start_seq_index:end_seq_index]
else:
retrieve_cum_len = None
if spec_info.seq_lens_cpu is not None: if spec_info.seq_lens_cpu is not None:
seq_lens_cpu = spec_info.seq_lens_cpu[start_seq_index:end_seq_index] seq_lens_cpu = spec_info.seq_lens_cpu[start_seq_index:end_seq_index]
@@ -252,10 +254,10 @@ def split_spec_info(
custom_mask=custom_mask, custom_mask=custom_mask,
draft_token=draft_token, draft_token=draft_token,
positions=positions, positions=positions,
retrive_index=retrive_index, retrieve_index=retrieve_index,
retrive_next_token=retrive_next_token, retrieve_next_token=retrieve_next_token,
retrive_next_sibling=retrive_next_sibling, retrieve_next_sibling=retrieve_next_sibling,
retrive_cum_len=retrive_cum_len, retrieve_cum_len=retrieve_cum_len,
seq_lens_cpu=seq_lens_cpu, seq_lens_cpu=seq_lens_cpu,
seq_lens_sum=seq_lens_sum, seq_lens_sum=seq_lens_sum,
) )
python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
+9 -7
View File
@@ -187,8 +187,10 @@ class MambaAttnBackendBase(AttentionBackend):
) )
if self.topk > 1: if self.topk > 1:
retrieve_next_token = forward_batch.spec_info.retrive_next_token retrieve_next_token = forward_batch.spec_info.retrieve_next_token
retrieve_next_sibling = forward_batch.spec_info.retrive_next_sibling retrieve_next_sibling = (
forward_batch.spec_info.retrieve_next_sibling
)
# retrieve_next_token is None during dummy run so skip tensor creation # retrieve_next_token is None during dummy run so skip tensor creation
if retrieve_next_token is not None: if retrieve_next_token is not None:
retrieve_parent_token = torch.empty_like(retrieve_next_token) retrieve_parent_token = torch.empty_like(retrieve_next_token)
@@ -485,8 +487,8 @@ class MambaAttnBackendBase(AttentionBackend):
# If topk > 1, we need to use retrieve_next_token and retrieve_next_sibling to handle the eagle tree custom attention mask # If topk > 1, we need to use retrieve_next_token and retrieve_next_sibling to handle the eagle tree custom attention mask
if forward_mode.is_target_verify() and self.topk > 1: if forward_mode.is_target_verify() and self.topk > 1:
# They are None during cuda graph capture so skip the copy_... # They are None during cuda graph capture so skip the copy_...
# self.retrieve_next_token_list[bs - 1].copy_(spec_info.retrive_next_token) # self.retrieve_next_token_list[bs - 1].copy_(spec_info.retrieve_next_token)
# self.retrieve_next_sibling_list[bs - 1].copy_(spec_info.retrive_next_sibling) # self.retrieve_next_sibling_list[bs - 1].copy_(spec_info.retrieve_next_sibling)
return ForwardMetadata( return ForwardMetadata(
query_start_loc=self.query_start_loc_list[bs - 1], query_start_loc=self.query_start_loc_list[bs - 1],
mamba_cache_indices=self.state_indices_list[bs - 1], mamba_cache_indices=self.state_indices_list[bs - 1],
@@ -545,12 +547,12 @@ class MambaAttnBackendBase(AttentionBackend):
# If topk > 1, we need to use retrieve_next_token and retrieve_next_sibling to handle the eagle tree custom attention mask # If topk > 1, we need to use retrieve_next_token and retrieve_next_sibling to handle the eagle tree custom attention mask
if forward_mode.is_target_verify() and self.topk > 1: if forward_mode.is_target_verify() and self.topk > 1:
bs_without_pad = spec_info.retrive_next_token.shape[0] bs_without_pad = spec_info.retrieve_next_token.shape[0]
self.retrieve_next_token_list[bs - 1][:bs_without_pad].copy_( self.retrieve_next_token_list[bs - 1][:bs_without_pad].copy_(
spec_info.retrive_next_token spec_info.retrieve_next_token
) )
self.retrieve_next_sibling_list[bs - 1][:bs_without_pad].copy_( self.retrieve_next_sibling_list[bs - 1][:bs_without_pad].copy_(
spec_info.retrive_next_sibling spec_info.retrieve_next_sibling
) )
return ForwardMetadata( return ForwardMetadata(
query_start_loc=self.query_start_loc_list[bs - 1], query_start_loc=self.query_start_loc_list[bs - 1],
python/sglang/srt/model_executor/cpu_graph_runner.py
+4 -4
View File
@@ -842,10 +842,10 @@ class CPUGraphRunner:
draft_token=None, draft_token=None,
custom_mask=self.custom_mask, custom_mask=self.custom_mask,
positions=None, positions=None,
retrive_index=None, retrieve_index=None,
retrive_next_token=None, retrieve_next_token=None,
retrive_next_sibling=None, retrieve_next_sibling=None,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.model_runner.server_args.speculative_num_steps, spec_steps=self.model_runner.server_args.speculative_num_steps,
topk=self.model_runner.server_args.speculative_eagle_topk, topk=self.model_runner.server_args.speculative_eagle_topk,
draft_token_num=self.model_runner.server_args.speculative_num_draft_tokens, draft_token_num=self.model_runner.server_args.speculative_num_draft_tokens,
python/sglang/srt/model_executor/cuda_graph_runner.py
+7 -7
View File
@@ -1334,10 +1334,10 @@ class CudaGraphRunner:
draft_token=None, draft_token=None,
custom_mask=self.buffers.custom_mask, custom_mask=self.buffers.custom_mask,
positions=None, positions=None,
retrive_index=None, retrieve_index=None,
retrive_next_token=None, retrieve_next_token=None,
retrive_next_sibling=None, retrieve_next_sibling=None,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.speculative_num_steps, spec_steps=self.speculative_num_steps,
topk=self.model_runner.server_args.speculative_eagle_topk, topk=self.model_runner.server_args.speculative_eagle_topk,
draft_token_num=self.speculative_num_draft_tokens, draft_token_num=self.speculative_num_draft_tokens,
@@ -1379,9 +1379,9 @@ class CudaGraphRunner:
draft_token=None, draft_token=None,
tree_mask=self.buffers.custom_mask, tree_mask=self.buffers.custom_mask,
positions=None, positions=None,
retrive_index=None, retrieve_index=None,
retrive_next_token=None, retrieve_next_token=None,
retrive_next_sibling=None, retrieve_next_sibling=None,
draft_token_num=self.num_tokens_per_bs, draft_token_num=self.num_tokens_per_bs,
) )
spec_info.capture_hidden_mode = CaptureHiddenMode.NULL spec_info.capture_hidden_mode = CaptureHiddenMode.NULL
python/sglang/srt/model_executor/model_runner.py
+7 -7
View File
@@ -2393,10 +2393,10 @@ class ModelRunner(ModelRunnerKVCacheMixin):
draft_token=None, draft_token=None,
custom_mask=buffers.custom_mask, custom_mask=buffers.custom_mask,
positions=None, positions=None,
retrive_index=None, retrieve_index=None,
retrive_next_token=None, retrieve_next_token=None,
retrive_next_sibling=None, retrieve_next_sibling=None,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.server_args.speculative_num_steps, spec_steps=self.server_args.speculative_num_steps,
topk=self.server_args.speculative_eagle_topk, topk=self.server_args.speculative_eagle_topk,
draft_token_num=self.server_args.speculative_num_draft_tokens, draft_token_num=self.server_args.speculative_num_draft_tokens,
@@ -2427,9 +2427,9 @@ class ModelRunner(ModelRunnerKVCacheMixin):
draft_token=None, draft_token=None,
tree_mask=buffers.custom_mask, tree_mask=buffers.custom_mask,
positions=None, positions=None,
retrive_index=None, retrieve_index=None,
retrive_next_token=None, retrieve_next_token=None,
retrive_next_sibling=None, retrieve_next_sibling=None,
draft_token_num=num_tokens_per_bs, draft_token_num=num_tokens_per_bs,
) )
spec_info.capture_hidden_mode = CaptureHiddenMode.NULL spec_info.capture_hidden_mode = CaptureHiddenMode.NULL
python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
+4 -4
View File
@@ -832,10 +832,10 @@ class PiecewiseCudaGraphRunner:
draft_token=None, draft_token=None,
custom_mask=self.custom_mask, custom_mask=self.custom_mask,
positions=None, positions=None,
retrive_index=None, retrieve_index=None,
retrive_next_token=None, retrieve_next_token=None,
retrive_next_sibling=None, retrieve_next_sibling=None,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.model_runner.server_args.speculative_num_steps, spec_steps=self.model_runner.server_args.speculative_num_steps,
topk=self.model_runner.server_args.speculative_eagle_topk, topk=self.model_runner.server_args.speculative_eagle_topk,
draft_token_num=self.model_runner.server_args.speculative_num_draft_tokens, draft_token_num=self.model_runner.server_args.speculative_num_draft_tokens,
python/sglang/srt/speculative/dflash_utils.py
+1
View File
@@ -618,6 +618,7 @@ def compute_dflash_sampling_accept_len_and_bonus(
accept_index=accept_index, accept_index=accept_index,
accept_token_num=accept_token_num, accept_token_num=accept_token_num,
candidates=candidates_i64, candidates=candidates_i64,
# kwarg LHS retained as `retrive_*` to match sgl_kernel op schema.
retrive_index=retrieve_index, retrive_index=retrieve_index,
retrive_next_token=retrieve_next_token, retrive_next_token=retrieve_next_token,
retrive_next_sibling=retrieve_next_sibling, retrive_next_sibling=retrieve_next_sibling,
python/sglang/srt/speculative/eagle_info.py
+20 -17
View File
@@ -61,10 +61,10 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
draft_token: torch.Tensor draft_token: torch.Tensor
custom_mask: torch.Tensor custom_mask: torch.Tensor
positions: torch.Tensor positions: torch.Tensor
retrive_index: torch.Tensor retrieve_index: torch.Tensor
retrive_next_token: torch.Tensor retrieve_next_token: torch.Tensor
retrive_next_sibling: torch.Tensor retrieve_next_sibling: torch.Tensor
retrive_cum_len: torch.Tensor retrieve_cum_len: torch.Tensor
spec_steps: int spec_steps: int
topk: int topk: int
draft_token_num: int draft_token_num: int
@@ -88,16 +88,16 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
draft_token=torch.empty((0,), dtype=torch.long, device="cuda"), draft_token=torch.empty((0,), dtype=torch.long, device="cuda"),
custom_mask=torch.full((0,), True, dtype=torch.bool, device="cuda"), custom_mask=torch.full((0,), True, dtype=torch.bool, device="cuda"),
positions=torch.empty((0,), dtype=torch.int64, device="cuda"), positions=torch.empty((0,), dtype=torch.int64, device="cuda"),
retrive_index=torch.full( retrieve_index=torch.full(
(0, num_verify_tokens), -1, dtype=torch.long, device="cuda" (0, num_verify_tokens), -1, dtype=torch.long, device="cuda"
), ),
retrive_next_token=torch.full( retrieve_next_token=torch.full(
(0, num_verify_tokens), -1, dtype=torch.long, device="cuda" (0, num_verify_tokens), -1, dtype=torch.long, device="cuda"
), ),
retrive_next_sibling=torch.full( retrieve_next_sibling=torch.full(
(0, num_verify_tokens), -1, dtype=torch.long, device="cuda" (0, num_verify_tokens), -1, dtype=torch.long, device="cuda"
), ),
retrive_cum_len=None, retrieve_cum_len=None,
topk=topk, topk=topk,
draft_token_num=num_verify_tokens, draft_token_num=num_verify_tokens,
spec_steps=spec_steps, spec_steps=spec_steps,
@@ -258,7 +258,7 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
), ),
) )
bs = self.retrive_index.shape[0] bs = self.retrieve_index.shape[0]
candidates = self.draft_token.reshape(bs, self.draft_token_num) candidates = self.draft_token.reshape(bs, self.draft_token_num)
sampling_info = batch.sampling_info sampling_info = batch.sampling_info
@@ -272,8 +272,10 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
if bs != len(sampling_info): if bs != len(sampling_info):
sampling_info = copy.deepcopy(sampling_info) sampling_info = copy.deepcopy(sampling_info)
# NOTE: retrive_index are the indices of the requests that are kept. # NOTE: retrieve_index are the indices of the requests that are kept.
sampling_info.filter_batch(self.retrive_index.tolist(), self.retrive_index) sampling_info.filter_batch(
self.retrieve_index.tolist(), self.retrieve_index
)
# Apply the custom logit processors if registered in the sampling info. # Apply the custom logit processors if registered in the sampling info.
if sampling_info.has_custom_logit_processor: if sampling_info.has_custom_logit_processor:
@@ -322,9 +324,9 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
accept_index=accept_index, # mutable accept_index=accept_index, # mutable
accept_token_num=accept_length, # mutable accept_token_num=accept_length, # mutable
candidates=candidates, candidates=candidates,
retrive_index=self.retrive_index, retrieve_index=self.retrieve_index,
retrive_next_token=self.retrive_next_token, retrieve_next_token=self.retrieve_next_token,
retrive_next_sibling=self.retrive_next_sibling, retrieve_next_sibling=self.retrieve_next_sibling,
target_predict=target_predict, target_predict=target_predict,
topk=self.topk, topk=self.topk,
) )
@@ -370,9 +372,10 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
accept_index=accept_index, # mutable accept_index=accept_index, # mutable
accept_token_num=accept_length, # mutable accept_token_num=accept_length, # mutable
candidates=candidates, candidates=candidates,
retrive_index=self.retrive_index, # kwarg LHS retained as `retrive_*` to match sgl_kernel op schema.
retrive_next_token=self.retrive_next_token, retrive_index=self.retrieve_index,
retrive_next_sibling=self.retrive_next_sibling, retrive_next_token=self.retrieve_next_token,
retrive_next_sibling=self.retrieve_next_sibling,
uniform_samples=coins, uniform_samples=coins,
uniform_samples_for_final_sampling=coins_for_final_sampling, uniform_samples_for_final_sampling=coins_for_final_sampling,
target_probs=target_probs, target_probs=target_probs,
python/sglang/srt/speculative/eagle_info_v2.py
+7 -6
View File
@@ -373,9 +373,9 @@ class EagleVerifyInputV2Mixin:
accept_index=accept_index, # mutable accept_index=accept_index, # mutable
accept_token_num=accept_length, # mutable accept_token_num=accept_length, # mutable
candidates=candidates, candidates=candidates,
retrive_index=self.retrive_index, retrieve_index=self.retrieve_index,
retrive_next_token=self.retrive_next_token, retrieve_next_token=self.retrieve_next_token,
retrive_next_sibling=self.retrive_next_sibling, retrieve_next_sibling=self.retrieve_next_sibling,
target_predict=target_predict, target_predict=target_predict,
topk=self.topk, topk=self.topk,
) )
@@ -415,9 +415,10 @@ class EagleVerifyInputV2Mixin:
accept_index=accept_index, # mutable accept_index=accept_index, # mutable
accept_token_num=accept_length, # mutable accept_token_num=accept_length, # mutable
candidates=candidates, candidates=candidates,
retrive_index=self.retrive_index, # kwarg LHS retained as `retrive_*` to match sgl_kernel op schema.
retrive_next_token=self.retrive_next_token, retrive_index=self.retrieve_index,
retrive_next_sibling=self.retrive_next_sibling, retrive_next_token=self.retrieve_next_token,
retrive_next_sibling=self.retrieve_next_sibling,
uniform_samples=coins, uniform_samples=coins,
uniform_samples_for_final_sampling=coins_for_final_sampling, uniform_samples_for_final_sampling=coins_for_final_sampling,
target_probs=target_probs, target_probs=target_probs,
python/sglang/srt/speculative/eagle_utils.py
+22 -20
View File
@@ -104,10 +104,10 @@ def build_tree_kernel_efficient(
raise NotImplementedError(f"Invalid tree mask: {tree_mask_mode=}") raise NotImplementedError(f"Invalid tree mask: {tree_mask_mode=}")
# TODO: make them torch.empty and fuse them into `sgl_build_tree_kernel` # TODO: make them torch.empty and fuse them into `sgl_build_tree_kernel`
retrive_buf = torch.full( retrieve_buf = torch.full(
(3, bs, num_verify_tokens), -1, device=device, dtype=torch.long (3, bs, num_verify_tokens), -1, device=device, dtype=torch.long
) )
retrive_index, retrive_next_token, retrive_next_sibling = retrive_buf retrieve_index, retrieve_next_token, retrieve_next_sibling = retrieve_buf
# position: where each token belongs to # position: where each token belongs to
# e.g. if depth of each draft token is [0, 1, 1, 2] and the prompt length is 7 # e.g. if depth of each draft token is [0, 1, 1, 2] and the prompt length is 7
# then, positions = [7, 8, 8, 9] # then, positions = [7, 8, 8, 9]
@@ -125,9 +125,9 @@ def build_tree_kernel_efficient(
seq_lens, seq_lens,
tree_mask, tree_mask,
positions, positions,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
topk, topk,
spec_steps, spec_steps,
num_verify_tokens, num_verify_tokens,
@@ -140,9 +140,9 @@ def build_tree_kernel_efficient(
seq_lens, seq_lens,
tree_mask, tree_mask,
positions, positions,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
topk, topk,
spec_steps, spec_steps,
num_verify_tokens, num_verify_tokens,
@@ -151,9 +151,9 @@ def build_tree_kernel_efficient(
return ( return (
tree_mask, tree_mask,
positions, positions,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
draft_tokens, draft_tokens,
) )
@@ -163,9 +163,9 @@ def verify_tree_greedy_func(
accept_index: torch.Tensor, accept_index: torch.Tensor,
accept_token_num: torch.Tensor, accept_token_num: torch.Tensor,
candidates: torch.Tensor, candidates: torch.Tensor,
retrive_index: torch.Tensor, retrieve_index: torch.Tensor,
retrive_next_token: torch.Tensor, retrieve_next_token: torch.Tensor,
retrive_next_sibling: torch.Tensor, retrieve_next_sibling: torch.Tensor,
target_predict: torch.Tensor, target_predict: torch.Tensor,
topk: int = -1, topk: int = -1,
): ):
@@ -177,9 +177,10 @@ def verify_tree_greedy_func(
accept_index=accept_index, # mutable accept_index=accept_index, # mutable
accept_token_num=accept_token_num, # mutable accept_token_num=accept_token_num, # mutable
candidates=candidates, candidates=candidates,
retrive_index=retrive_index, # kwarg LHS retained as `retrive_*` to match sgl_kernel op schema.
retrive_next_token=retrive_next_token, retrive_index=retrieve_index,
retrive_next_sibling=retrive_next_sibling, retrive_next_token=retrieve_next_token,
retrive_next_sibling=retrieve_next_sibling,
target_predict=target_predict, target_predict=target_predict,
) )
@@ -191,9 +192,10 @@ def verify_tree_greedy_func(
accept_index=accept_index, accept_index=accept_index,
accept_token_num=accept_token_num, accept_token_num=accept_token_num,
candidates=candidates, candidates=candidates,
retrive_index=retrive_index, # kwarg LHS retained as `retrive_*` to match sgl_kernel op schema.
retrive_next_token=retrive_next_token, retrive_index=retrieve_index,
retrive_next_sibling=retrive_next_sibling, retrive_next_token=retrieve_next_token,
retrive_next_sibling=retrieve_next_sibling,
target_predict=target_predict, target_predict=target_predict,
) )
return predicts, accept_index, accept_token_num return predicts, accept_index, accept_token_num
python/sglang/srt/speculative/eagle_worker.py
+11 -11
View File
@@ -763,9 +763,9 @@ class EAGLEWorker(TpModelWorker):
( (
tree_mask, tree_mask,
position, position,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
draft_tokens, draft_tokens,
) = build_tree_kernel_efficient( ) = build_tree_kernel_efficient(
spec_info.verified_id, spec_info.verified_id,
@@ -783,10 +783,10 @@ class EAGLEWorker(TpModelWorker):
draft_token=draft_tokens, draft_token=draft_tokens,
custom_mask=tree_mask, custom_mask=tree_mask,
positions=position, positions=position,
retrive_index=retrive_index, retrieve_index=retrieve_index,
retrive_next_token=retrive_next_token, retrieve_next_token=retrieve_next_token,
retrive_next_sibling=retrive_next_sibling, retrieve_next_sibling=retrieve_next_sibling,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.speculative_num_steps, spec_steps=self.speculative_num_steps,
topk=self.topk, topk=self.topk,
draft_token_num=self.speculative_num_draft_tokens, draft_token_num=self.speculative_num_draft_tokens,
@@ -897,10 +897,10 @@ class EAGLEWorker(TpModelWorker):
assert model_worker_batch.capture_hidden_mode == spec_info.capture_hidden_mode assert model_worker_batch.capture_hidden_mode == spec_info.capture_hidden_mode
if batch.has_grammar: if batch.has_grammar:
retrieve_next_token_cpu = spec_info.retrive_next_token.cpu() retrieve_next_token_cpu = spec_info.retrieve_next_token.cpu()
retrieve_next_sibling_cpu = spec_info.retrive_next_sibling.cpu() retrieve_next_sibling_cpu = spec_info.retrieve_next_sibling.cpu()
draft_tokens_cpu = spec_info.draft_token.view( draft_tokens_cpu = spec_info.draft_token.view(
spec_info.retrive_next_token.shape spec_info.retrieve_next_token.shape
).cpu() ).cpu()
# Forward # Forward
@@ -927,7 +927,7 @@ class EAGLEWorker(TpModelWorker):
if vocab_mask is not None: if vocab_mask is not None:
assert spec_info.grammar is not None assert spec_info.grammar is not None
vocab_mask = vocab_mask.to(spec_info.retrive_next_token.device) vocab_mask = vocab_mask.to(spec_info.retrieve_next_token.device)
# NOTE (sk): otherwise, this vocab mask will be the one from the previous extend stage # NOTE (sk): otherwise, this vocab mask will be the one from the previous extend stage
# and will be applied to produce wrong results # and will be applied to produce wrong results
batch.sampling_info.vocab_mask = None batch.sampling_info.vocab_mask = None
python/sglang/srt/speculative/eagle_worker_v2.py
+11 -11
View File
@@ -364,9 +364,9 @@ class EagleDraftWorker(BaseDraftWorker):
( (
tree_mask, tree_mask,
position, position,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
draft_tokens, draft_tokens,
) = build_tree_kernel_efficient( ) = build_tree_kernel_efficient(
draft_input.verified_id, draft_input.verified_id,
@@ -387,10 +387,10 @@ class EagleDraftWorker(BaseDraftWorker):
draft_token=draft_tokens, draft_token=draft_tokens,
custom_mask=tree_mask, custom_mask=tree_mask,
positions=position, positions=position,
retrive_index=retrive_index, retrieve_index=retrieve_index,
retrive_next_token=retrive_next_token, retrieve_next_token=retrieve_next_token,
retrive_next_sibling=retrive_next_sibling, retrieve_next_sibling=retrieve_next_sibling,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.speculative_num_steps, spec_steps=self.speculative_num_steps,
topk=self.topk, topk=self.topk,
draft_token_num=self.speculative_num_draft_tokens, draft_token_num=self.speculative_num_draft_tokens,
@@ -786,10 +786,10 @@ class EAGLEWorkerV2(BaseSpecWorker):
# Prepare grammar data on CPU if needed # Prepare grammar data on CPU if needed
if batch.has_grammar: if batch.has_grammar:
retrieve_next_token_cpu = verify_input.retrive_next_token.cpu() retrieve_next_token_cpu = verify_input.retrieve_next_token.cpu()
retrieve_next_sibling_cpu = verify_input.retrive_next_sibling.cpu() retrieve_next_sibling_cpu = verify_input.retrieve_next_sibling.cpu()
draft_tokens_cpu = verify_input.draft_token.view( draft_tokens_cpu = verify_input.draft_token.view(
verify_input.retrive_next_token.shape verify_input.retrieve_next_token.shape
).cpu() ).cpu()
# Run target verify batch in the main compute stream (GPU compute) # Run target verify batch in the main compute stream (GPU compute)
@@ -816,7 +816,7 @@ class EAGLEWorkerV2(BaseSpecWorker):
if vocab_mask is not None: if vocab_mask is not None:
assert verify_input.grammar is not None assert verify_input.grammar is not None
vocab_mask = vocab_mask.to(verify_input.retrive_next_token.device) vocab_mask = vocab_mask.to(verify_input.retrieve_next_token.device)
# NOTE: otherwise, this vocab mask will be the one from the previous extend stage # NOTE: otherwise, this vocab mask will be the one from the previous extend stage
# and will be applied to produce wrong results # and will be applied to produce wrong results
batch.sampling_info.vocab_mask = None batch.sampling_info.vocab_mask = None
python/sglang/srt/speculative/multi_layer_eagle_worker.py
+11 -11
View File
@@ -435,9 +435,9 @@ class MultiLayerEagleWorker(TpModelWorker):
( (
tree_mask, tree_mask,
position, position,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
draft_tokens, draft_tokens,
) = build_tree_kernel_efficient( ) = build_tree_kernel_efficient(
spec_info.verified_id, spec_info.verified_id,
@@ -455,10 +455,10 @@ class MultiLayerEagleWorker(TpModelWorker):
draft_token=draft_tokens, draft_token=draft_tokens,
custom_mask=tree_mask, custom_mask=tree_mask,
positions=position, positions=position,
retrive_index=retrive_index, retrieve_index=retrieve_index,
retrive_next_token=retrive_next_token, retrieve_next_token=retrieve_next_token,
retrive_next_sibling=retrive_next_sibling, retrieve_next_sibling=retrieve_next_sibling,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.speculative_num_steps, spec_steps=self.speculative_num_steps,
topk=self.topk, topk=self.topk,
draft_token_num=self.server_args.speculative_num_draft_tokens, draft_token_num=self.server_args.speculative_num_draft_tokens,
@@ -488,10 +488,10 @@ class MultiLayerEagleWorker(TpModelWorker):
model_worker_batch.return_hidden_states_before_norm = True model_worker_batch.return_hidden_states_before_norm = True
if batch.has_grammar: if batch.has_grammar:
retrieve_next_token_cpu = spec_info.retrive_next_token.cpu() retrieve_next_token_cpu = spec_info.retrieve_next_token.cpu()
retrieve_next_sibling_cpu = spec_info.retrive_next_sibling.cpu() retrieve_next_sibling_cpu = spec_info.retrieve_next_sibling.cpu()
draft_tokens_cpu = spec_info.draft_token.view( draft_tokens_cpu = spec_info.draft_token.view(
spec_info.retrive_next_token.shape spec_info.retrieve_next_token.shape
).cpu() ).cpu()
# Forward # Forward
@@ -518,7 +518,7 @@ class MultiLayerEagleWorker(TpModelWorker):
if vocab_mask is not None: if vocab_mask is not None:
assert spec_info.grammar is not None assert spec_info.grammar is not None
vocab_mask = vocab_mask.to(spec_info.retrive_next_token.device) vocab_mask = vocab_mask.to(spec_info.retrieve_next_token.device)
# NOTE (sk): otherwise, this vocab mask will be the one from the previous extend stage # NOTE (sk): otherwise, this vocab mask will be the one from the previous extend stage
# and will be applied to produce wrong results # and will be applied to produce wrong results
batch.sampling_info.vocab_mask = None batch.sampling_info.vocab_mask = None
python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py
+7 -7
View File
@@ -248,9 +248,9 @@ class MultiLayerEagleDraftWorker(BaseDraftWorker):
( (
tree_mask, tree_mask,
position, position,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
draft_tokens, draft_tokens,
) = build_tree_kernel_efficient( ) = build_tree_kernel_efficient(
draft_input.verified_id, draft_input.verified_id,
@@ -271,10 +271,10 @@ class MultiLayerEagleDraftWorker(BaseDraftWorker):
draft_token=draft_tokens, draft_token=draft_tokens,
custom_mask=tree_mask, custom_mask=tree_mask,
positions=position, positions=position,
retrive_index=retrive_index, retrieve_index=retrieve_index,
retrive_next_token=retrive_next_token, retrieve_next_token=retrieve_next_token,
retrive_next_sibling=retrive_next_sibling, retrieve_next_sibling=retrieve_next_sibling,
retrive_cum_len=None, retrieve_cum_len=None,
spec_steps=self.speculative_num_steps, spec_steps=self.speculative_num_steps,
topk=self.topk, topk=self.topk,
draft_token_num=self.speculative_num_draft_tokens, draft_token_num=self.speculative_num_draft_tokens,
python/sglang/srt/speculative/ngram_info.py
+18 -15
View File
@@ -54,9 +54,9 @@ class NgramVerifyInput(SpecInput):
draft_token: torch.Tensor, draft_token: torch.Tensor,
tree_mask: torch.Tensor, tree_mask: torch.Tensor,
positions: torch.Tensor, positions: torch.Tensor,
retrive_index: torch.Tensor, retrieve_index: torch.Tensor,
retrive_next_token: torch.Tensor, retrieve_next_token: torch.Tensor,
retrive_next_sibling: torch.Tensor, retrieve_next_sibling: torch.Tensor,
draft_token_num: int, draft_token_num: int,
grammar: BaseGrammarObject = None, grammar: BaseGrammarObject = None,
): ):
@@ -64,9 +64,9 @@ class NgramVerifyInput(SpecInput):
self.draft_token = draft_token self.draft_token = draft_token
self.custom_mask = tree_mask self.custom_mask = tree_mask
self.positions = positions self.positions = positions
self.retrive_index = retrive_index self.retrieve_index = retrieve_index
self.retrive_next_token = retrive_next_token self.retrieve_next_token = retrieve_next_token
self.retrive_next_sibling = retrive_next_sibling self.retrieve_next_sibling = retrieve_next_sibling
self.draft_token_num = draft_token_num self.draft_token_num = draft_token_num
self.device = self.custom_mask.device self.device = self.custom_mask.device
self.grammar = grammar self.grammar = grammar
@@ -303,9 +303,9 @@ class NgramVerifyInput(SpecInput):
accept_index=self.accepted_indices, # mutable accept_index=self.accepted_indices, # mutable
accept_token_num=self.accept_length, # mutable accept_token_num=self.accept_length, # mutable
candidates=candidates, candidates=candidates,
retrive_index=self.retrive_index, retrieve_index=self.retrieve_index,
retrive_next_token=self.retrive_next_token, retrieve_next_token=self.retrieve_next_token,
retrive_next_sibling=self.retrive_next_sibling, retrieve_next_sibling=self.retrieve_next_sibling,
target_predict=target_predict, target_predict=target_predict,
) )
@@ -365,9 +365,10 @@ class NgramVerifyInput(SpecInput):
accept_index=self.accepted_indices, # mutable accept_index=self.accepted_indices, # mutable
accept_token_num=self.accept_length, # mutable accept_token_num=self.accept_length, # mutable
candidates=candidates.to(torch.int64), candidates=candidates.to(torch.int64),
retrive_index=self.retrive_index.to(torch.int64), # kwarg LHS retained as `retrive_*` to match sgl_kernel op schema.
retrive_next_token=self.retrive_next_token.to(torch.int64), retrive_index=self.retrieve_index.to(torch.int64),
retrive_next_sibling=self.retrive_next_sibling.to(torch.int64), retrive_next_token=self.retrieve_next_token.to(torch.int64),
retrive_next_sibling=self.retrieve_next_sibling.to(torch.int64),
uniform_samples=coins, uniform_samples=coins,
uniform_samples_for_final_sampling=coins_for_final_sampling, uniform_samples_for_final_sampling=coins_for_final_sampling,
target_probs=target_probs, target_probs=target_probs,
@@ -384,13 +385,15 @@ class NgramVerifyInput(SpecInput):
page_size: int, page_size: int,
vocab_mask: Optional[torch.Tensor] = None, # For grammar vocab_mask: Optional[torch.Tensor] = None, # For grammar
) -> torch.Tensor: ) -> torch.Tensor:
bs = self.retrive_index.shape[0] bs = self.retrieve_index.shape[0]
sampling_info = batch.sampling_info sampling_info = batch.sampling_info
if bs != len(sampling_info): if bs != len(sampling_info):
sampling_info = copy.deepcopy(sampling_info) sampling_info = copy.deepcopy(sampling_info)
# NOTE: retrive_index are the indices of the requests that are kept. # NOTE: retrieve_index are the indices of the requests that are kept.
sampling_info.filter_batch(self.retrive_index.tolist(), self.retrive_index) sampling_info.filter_batch(
self.retrieve_index.tolist(), self.retrieve_index
)
# Apply the custom logit processors if registered in the sampling info. # Apply the custom logit processors if registered in the sampling info.
if sampling_info.has_custom_logit_processor: if sampling_info.has_custom_logit_processor:
python/sglang/srt/speculative/ngram_worker.py
+19 -19
View File
@@ -117,12 +117,12 @@ class NGRAMWorker:
dtype=torch.int64, dtype=torch.int64,
device=self.device, device=self.device,
) )
self.retrive_next_token = torch.empty( self.retrieve_next_token = torch.empty(
(self.max_batch_size, self.draft_token_num), (self.max_batch_size, self.draft_token_num),
dtype=torch.int64, dtype=torch.int64,
device=self.device, device=self.device,
) )
self.retrive_next_sibling = torch.empty( self.retrieve_next_sibling = torch.empty(
(self.max_batch_size, self.draft_token_num), (self.max_batch_size, self.draft_token_num),
dtype=torch.int64, dtype=torch.int64,
device=self.device, device=self.device,
@@ -137,14 +137,14 @@ class NGRAMWorker:
self.draft_tokens_batch = [] self.draft_tokens_batch = []
self.tree_mask_batch = [] self.tree_mask_batch = []
self.retrieve_indexes_batch = [] self.retrieve_indexes_batch = []
self.retrive_next_token_batch = [] self.retrieve_next_token_batch = []
self.retrive_next_sibling_batch = [] self.retrieve_next_sibling_batch = []
self.positions_batch = [] self.positions_batch = []
for bs in range(0, self.max_batch_size + 1): for bs in range(0, self.max_batch_size + 1):
self.retrieve_indexes_batch.append(self.retrieve_indexes[:bs, :]) self.retrieve_indexes_batch.append(self.retrieve_indexes[:bs, :])
self.retrive_next_token_batch.append(self.retrive_next_token[:bs, :]) self.retrieve_next_token_batch.append(self.retrieve_next_token[:bs, :])
self.retrive_next_sibling_batch.append(self.retrive_next_sibling[:bs, :]) self.retrieve_next_sibling_batch.append(self.retrieve_next_sibling[:bs, :])
self.positions_batch.append(self.positions[: bs * self.draft_token_num]) self.positions_batch.append(self.positions[: bs * self.draft_token_num])
self.draft_tokens_batch.append( self.draft_tokens_batch.append(
self.draft_tokens[: bs * self.draft_token_num] self.draft_tokens[: bs * self.draft_token_num]
@@ -186,9 +186,9 @@ class NGRAMWorker:
bs = batch.batch_size() bs = batch.batch_size()
retrive_index = self.retrieve_indexes_batch[bs] retrieve_index = self.retrieve_indexes_batch[bs]
retrive_next_token = self.retrive_next_token_batch[bs] retrieve_next_token = self.retrieve_next_token_batch[bs]
retrive_next_sibling = self.retrive_next_sibling_batch[bs] retrieve_next_sibling = self.retrieve_next_sibling_batch[bs]
positions = self.positions_batch[bs] positions = self.positions_batch[bs]
tree_mask = self.tree_mask_batch[bs] tree_mask = self.tree_mask_batch[bs]
draft_tokens = self.draft_tokens_batch[bs] draft_tokens = self.draft_tokens_batch[bs]
@@ -201,9 +201,9 @@ class NGRAMWorker:
tree_mask, tree_mask,
batch.seq_lens, batch.seq_lens,
positions, # mutable positions, # mutable
retrive_index, # mutable retrieve_index, # mutable
retrive_next_token, # mutable retrieve_next_token, # mutable
retrive_next_sibling, # mutable retrieve_next_sibling, # mutable
bs, bs,
self.draft_token_num, self.draft_token_num,
) )
@@ -230,9 +230,9 @@ class NGRAMWorker:
draft_tokens, draft_tokens,
tree_mask, tree_mask,
positions, positions,
retrive_index, retrieve_index,
retrive_next_token, retrieve_next_token,
retrive_next_sibling, retrieve_next_sibling,
self.draft_token_num, self.draft_token_num,
) )
batch.spec_info.prepare_for_verify(batch, self.page_size) batch.spec_info.prepare_for_verify(batch, self.page_size)
@@ -266,10 +266,10 @@ class NGRAMWorker:
if model_worker_batch.forward_mode.is_target_verify(): if model_worker_batch.forward_mode.is_target_verify():
if batch.has_grammar: if batch.has_grammar:
retrieve_next_token_cpu = spec_info.retrive_next_token.cpu() retrieve_next_token_cpu = spec_info.retrieve_next_token.cpu()
retrieve_next_sibling_cpu = spec_info.retrive_next_sibling.cpu() retrieve_next_sibling_cpu = spec_info.retrieve_next_sibling.cpu()
draft_tokens_cpu = spec_info.draft_token.view( draft_tokens_cpu = spec_info.draft_token.view(
spec_info.retrive_next_token.shape spec_info.retrieve_next_token.shape
).cpu() ).cpu()
set_time_batch(batch.reqs, "set_spec_verify_start_time", trace_only=True) set_time_batch(batch.reqs, "set_spec_verify_start_time", trace_only=True)
@@ -298,7 +298,7 @@ class NGRAMWorker:
if vocab_mask is not None: if vocab_mask is not None:
assert verify_input.grammar is not None assert verify_input.grammar is not None
vocab_mask = vocab_mask.to(verify_input.retrive_next_token.device) vocab_mask = vocab_mask.to(verify_input.retrieve_next_token.device)
# NOTE (sk): otherwise, this vocab mask will be the one from the previous extend stage # NOTE (sk): otherwise, this vocab mask will be the one from the previous extend stage
# and will be applied to produce wrong results # and will be applied to produce wrong results
batch.sampling_info.vocab_mask = None batch.sampling_info.vocab_mask = None