sglang 0.5.3rc0__py3-none-any.whl → 0.5.3rc2__py3-none-any.whl

sglang/srt/grpc/sglang_scheduler_pb2.pyi

@@ -3,7 +3,6 @@ import datetime
 from google.protobuf import timestamp_pb2 as _timestamp_pb2
 from google.protobuf import struct_pb2 as _struct_pb2
 from google.protobuf.internal import containers as _containers
-from google.protobuf.internal import enum_type_wrapper as _enum_type_wrapper
 from google.protobuf import descriptor as _descriptor
 from google.protobuf import message as _message
 from collections.abc import Iterable as _Iterable, Mapping as _Mapping
@@ -12,7 +11,7 @@ from typing import ClassVar as _ClassVar, Optional as _Optional, Union as _Union
 DESCRIPTOR: _descriptor.FileDescriptor
 
 class SamplingParams(_message.Message):
-    __slots__ = ("temperature", "top_p", "top_k", "min_p", "frequency_penalty", "presence_penalty", "repetition_penalty", "max_new_tokens", "stop", "stop_token_ids", "skip_special_tokens", "spaces_between_special_tokens", "regex", "json_schema", "ebnf_grammar", "lora_path", "n", "token_healing", "min_new_tokens", "ignore_eos", "no_stop_trim", "stream_interval", "logit_bias", "structural_tag", "custom_params")
+    __slots__ = ("temperature", "top_p", "top_k", "min_p", "frequency_penalty", "presence_penalty", "repetition_penalty", "max_new_tokens", "stop", "stop_token_ids", "skip_special_tokens", "spaces_between_special_tokens", "regex", "json_schema", "ebnf_grammar", "structural_tag", "lora_path", "n", "token_healing", "min_new_tokens", "ignore_eos", "no_stop_trim", "stream_interval", "logit_bias", "custom_params")
     class LogitBiasEntry(_message.Message):
         __slots__ = ("key", "value")
         KEY_FIELD_NUMBER: _ClassVar[int]
@@ -35,6 +34,7 @@ class SamplingParams(_message.Message):
     REGEX_FIELD_NUMBER: _ClassVar[int]
     JSON_SCHEMA_FIELD_NUMBER: _ClassVar[int]
     EBNF_GRAMMAR_FIELD_NUMBER: _ClassVar[int]
+    STRUCTURAL_TAG_FIELD_NUMBER: _ClassVar[int]
     LORA_PATH_FIELD_NUMBER: _ClassVar[int]
     N_FIELD_NUMBER: _ClassVar[int]
     TOKEN_HEALING_FIELD_NUMBER: _ClassVar[int]
@@ -43,7 +43,6 @@ class SamplingParams(_message.Message):
     NO_STOP_TRIM_FIELD_NUMBER: _ClassVar[int]
     STREAM_INTERVAL_FIELD_NUMBER: _ClassVar[int]
     LOGIT_BIAS_FIELD_NUMBER: _ClassVar[int]
-    STRUCTURAL_TAG_FIELD_NUMBER: _ClassVar[int]
     CUSTOM_PARAMS_FIELD_NUMBER: _ClassVar[int]
     temperature: float
     top_p: float
@@ -60,6 +59,7 @@ class SamplingParams(_message.Message):
     regex: str
     json_schema: str
     ebnf_grammar: str
+    structural_tag: str
     lora_path: str
     n: int
     token_healing: bool
@@ -68,9 +68,8 @@ class SamplingParams(_message.Message):
     no_stop_trim: bool
     stream_interval: int
     logit_bias: _containers.ScalarMap[str, float]
-    structural_tag: str
     custom_params: _struct_pb2.Struct
-    def __init__(self, temperature: _Optional[float] = ..., top_p: _Optional[float] = ..., top_k: _Optional[int] = ..., min_p: _Optional[float] = ..., frequency_penalty: _Optional[float] = ..., presence_penalty: _Optional[float] = ..., repetition_penalty: _Optional[float] = ..., max_new_tokens: _Optional[int] = ..., stop: _Optional[_Iterable[str]] = ..., stop_token_ids: _Optional[_Iterable[int]] = ..., skip_special_tokens: bool = ..., spaces_between_special_tokens: bool = ..., regex: _Optional[str] = ..., json_schema: _Optional[str] = ..., ebnf_grammar: _Optional[str] = ..., lora_path: _Optional[str] = ..., n: _Optional[int] = ..., token_healing: bool = ..., min_new_tokens: _Optional[int] = ..., ignore_eos: bool = ..., no_stop_trim: bool = ..., stream_interval: _Optional[int] = ..., logit_bias: _Optional[_Mapping[str, float]] = ..., structural_tag: _Optional[str] = ..., custom_params: _Optional[_Union[_struct_pb2.Struct, _Mapping]] = ...) -> None: ...
+    def __init__(self, temperature: _Optional[float] = ..., top_p: _Optional[float] = ..., top_k: _Optional[int] = ..., min_p: _Optional[float] = ..., frequency_penalty: _Optional[float] = ..., presence_penalty: _Optional[float] = ..., repetition_penalty: _Optional[float] = ..., max_new_tokens: _Optional[int] = ..., stop: _Optional[_Iterable[str]] = ..., stop_token_ids: _Optional[_Iterable[int]] = ..., skip_special_tokens: bool = ..., spaces_between_special_tokens: bool = ..., regex: _Optional[str] = ..., json_schema: _Optional[str] = ..., ebnf_grammar: _Optional[str] = ..., structural_tag: _Optional[str] = ..., lora_path: _Optional[str] = ..., n: _Optional[int] = ..., token_healing: bool = ..., min_new_tokens: _Optional[int] = ..., ignore_eos: bool = ..., no_stop_trim: bool = ..., stream_interval: _Optional[int] = ..., logit_bias: _Optional[_Mapping[str, float]] = ..., custom_params: _Optional[_Union[_struct_pb2.Struct, _Mapping]] = ...) -> None: ...
 
 class DisaggregatedParams(_message.Message):
     __slots__ = ("bootstrap_host", "bootstrap_port", "bootstrap_room")
@@ -83,7 +82,7 @@ class DisaggregatedParams(_message.Message):
     def __init__(self, bootstrap_host: _Optional[str] = ..., bootstrap_port: _Optional[int] = ..., bootstrap_room: _Optional[int] = ...) -> None: ...
 
 class GenerateRequest(_message.Message):
-    __slots__ = ("request_id", "tokenized", "mm_inputs", "sampling_params", "return_logprob", "logprob_start_len", "top_logprobs_num", "token_ids_logprob", "return_hidden_states", "disaggregated_params", "custom_logit_processor", "timestamp", "log_metrics", "input_embeds", "lora_id", "data_parallel_rank", "dp_balance_id")
+    __slots__ = ("request_id", "tokenized", "mm_inputs", "sampling_params", "return_logprob", "logprob_start_len", "top_logprobs_num", "token_ids_logprob", "return_hidden_states", "disaggregated_params", "custom_logit_processor", "timestamp", "log_metrics", "input_embeds", "lora_id", "data_parallel_rank", "stream")
     REQUEST_ID_FIELD_NUMBER: _ClassVar[int]
     TOKENIZED_FIELD_NUMBER: _ClassVar[int]
     MM_INPUTS_FIELD_NUMBER: _ClassVar[int]
@@ -100,7 +99,7 @@ class GenerateRequest(_message.Message):
     INPUT_EMBEDS_FIELD_NUMBER: _ClassVar[int]
     LORA_ID_FIELD_NUMBER: _ClassVar[int]
     DATA_PARALLEL_RANK_FIELD_NUMBER: _ClassVar[int]
-    DP_BALANCE_ID_FIELD_NUMBER: _ClassVar[int]
+    STREAM_FIELD_NUMBER: _ClassVar[int]
     request_id: str
     tokenized: TokenizedInput
     mm_inputs: MultimodalInputs
@@ -117,8 +116,8 @@ class GenerateRequest(_message.Message):
     input_embeds: _containers.RepeatedScalarFieldContainer[float]
     lora_id: str
     data_parallel_rank: int
-    dp_balance_id: int
-    def __init__(self, request_id: _Optional[str] = ..., tokenized: _Optional[_Union[TokenizedInput, _Mapping]] = ..., mm_inputs: _Optional[_Union[MultimodalInputs, _Mapping]] = ..., sampling_params: _Optional[_Union[SamplingParams, _Mapping]] = ..., return_logprob: bool = ..., logprob_start_len: _Optional[int] = ..., top_logprobs_num: _Optional[int] = ..., token_ids_logprob: _Optional[_Iterable[int]] = ..., return_hidden_states: bool = ..., disaggregated_params: _Optional[_Union[DisaggregatedParams, _Mapping]] = ..., custom_logit_processor: _Optional[str] = ..., timestamp: _Optional[_Union[datetime.datetime, _timestamp_pb2.Timestamp, _Mapping]] = ..., log_metrics: bool = ..., input_embeds: _Optional[_Iterable[float]] = ..., lora_id: _Optional[str] = ..., data_parallel_rank: _Optional[int] = ..., dp_balance_id: _Optional[int] = ...) -> None: ...
+    stream: bool
+    def __init__(self, request_id: _Optional[str] = ..., tokenized: _Optional[_Union[TokenizedInput, _Mapping]] = ..., mm_inputs: _Optional[_Union[MultimodalInputs, _Mapping]] = ..., sampling_params: _Optional[_Union[SamplingParams, _Mapping]] = ..., return_logprob: bool = ..., logprob_start_len: _Optional[int] = ..., top_logprobs_num: _Optional[int] = ..., token_ids_logprob: _Optional[_Iterable[int]] = ..., return_hidden_states: bool = ..., disaggregated_params: _Optional[_Union[DisaggregatedParams, _Mapping]] = ..., custom_logit_processor: _Optional[str] = ..., timestamp: _Optional[_Union[datetime.datetime, _timestamp_pb2.Timestamp, _Mapping]] = ..., log_metrics: bool = ..., input_embeds: _Optional[_Iterable[float]] = ..., lora_id: _Optional[str] = ..., data_parallel_rank: _Optional[int] = ..., stream: bool = ...) -> None: ...
 
 class TokenizedInput(_message.Message):
     __slots__ = ("original_text", "input_ids")
@@ -161,52 +160,50 @@ class GenerateResponse(_message.Message):
     def __init__(self, request_id: _Optional[str] = ..., chunk: _Optional[_Union[GenerateStreamChunk, _Mapping]] = ..., complete: _Optional[_Union[GenerateComplete, _Mapping]] = ..., error: _Optional[_Union[GenerateError, _Mapping]] = ...) -> None: ...
 
 class GenerateStreamChunk(_message.Message):
-    __slots__ = ("token_id", "text", "prompt_tokens", "completion_tokens", "cached_tokens", "logprobs", "hidden_states", "generation_time", "queue_time")
-    TOKEN_ID_FIELD_NUMBER: _ClassVar[int]
-    TEXT_FIELD_NUMBER: _ClassVar[int]
+    __slots__ = ("token_ids", "prompt_tokens", "completion_tokens", "cached_tokens", "output_logprobs", "hidden_states", "input_logprobs", "index")
+    TOKEN_IDS_FIELD_NUMBER: _ClassVar[int]
     PROMPT_TOKENS_FIELD_NUMBER: _ClassVar[int]
     COMPLETION_TOKENS_FIELD_NUMBER: _ClassVar[int]
     CACHED_TOKENS_FIELD_NUMBER: _ClassVar[int]
-    LOGPROBS_FIELD_NUMBER: _ClassVar[int]
+    OUTPUT_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
     HIDDEN_STATES_FIELD_NUMBER: _ClassVar[int]
-    GENERATION_TIME_FIELD_NUMBER: _ClassVar[int]
-    QUEUE_TIME_FIELD_NUMBER: _ClassVar[int]
-    token_id: int
-    text: str
+    INPUT_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
+    INDEX_FIELD_NUMBER: _ClassVar[int]
+    token_ids: _containers.RepeatedScalarFieldContainer[int]
     prompt_tokens: int
     completion_tokens: int
     cached_tokens: int
-    logprobs: LogProbs
+    output_logprobs: OutputLogProbs
     hidden_states: _containers.RepeatedScalarFieldContainer[float]
-    generation_time: float
-    queue_time: int
-    def __init__(self, token_id: _Optional[int] = ..., text: _Optional[str] = ..., prompt_tokens: _Optional[int] = ..., completion_tokens: _Optional[int] = ..., cached_tokens: _Optional[int] = ..., logprobs: _Optional[_Union[LogProbs, _Mapping]] = ..., hidden_states: _Optional[_Iterable[float]] = ..., generation_time: _Optional[float] = ..., queue_time: _Optional[int] = ...) -> None: ...
+    input_logprobs: InputLogProbs
+    index: int
+    def __init__(self, token_ids: _Optional[_Iterable[int]] = ..., prompt_tokens: _Optional[int] = ..., completion_tokens: _Optional[int] = ..., cached_tokens: _Optional[int] = ..., output_logprobs: _Optional[_Union[OutputLogProbs, _Mapping]] = ..., hidden_states: _Optional[_Iterable[float]] = ..., input_logprobs: _Optional[_Union[InputLogProbs, _Mapping]] = ..., index: _Optional[int] = ...) -> None: ...
 
 class GenerateComplete(_message.Message):
-    __slots__ = ("output_ids", "output_text", "finish_reason", "all_logprobs", "all_hidden_states")
-    class FinishReason(int, metaclass=_enum_type_wrapper.EnumTypeWrapper):
-        __slots__ = ()
-        STOP: _ClassVar[GenerateComplete.FinishReason]
-        LENGTH: _ClassVar[GenerateComplete.FinishReason]
-        EOS_TOKEN: _ClassVar[GenerateComplete.FinishReason]
-        STOP_STR: _ClassVar[GenerateComplete.FinishReason]
-        ABORT: _ClassVar[GenerateComplete.FinishReason]
-    STOP: GenerateComplete.FinishReason
-    LENGTH: GenerateComplete.FinishReason
-    EOS_TOKEN: GenerateComplete.FinishReason
-    STOP_STR: GenerateComplete.FinishReason
-    ABORT: GenerateComplete.FinishReason
+    __slots__ = ("output_ids", "finish_reason", "prompt_tokens", "completion_tokens", "cached_tokens", "output_logprobs", "all_hidden_states", "matched_token_id", "matched_stop_str", "input_logprobs", "index")
     OUTPUT_IDS_FIELD_NUMBER: _ClassVar[int]
-    OUTPUT_TEXT_FIELD_NUMBER: _ClassVar[int]
     FINISH_REASON_FIELD_NUMBER: _ClassVar[int]
-    ALL_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
+    PROMPT_TOKENS_FIELD_NUMBER: _ClassVar[int]
+    COMPLETION_TOKENS_FIELD_NUMBER: _ClassVar[int]
+    CACHED_TOKENS_FIELD_NUMBER: _ClassVar[int]
+    OUTPUT_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
     ALL_HIDDEN_STATES_FIELD_NUMBER: _ClassVar[int]
+    MATCHED_TOKEN_ID_FIELD_NUMBER: _ClassVar[int]
+    MATCHED_STOP_STR_FIELD_NUMBER: _ClassVar[int]
+    INPUT_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
+    INDEX_FIELD_NUMBER: _ClassVar[int]
     output_ids: _containers.RepeatedScalarFieldContainer[int]
-    output_text: str
-    finish_reason: GenerateComplete.FinishReason
-    all_logprobs: _containers.RepeatedCompositeFieldContainer[LogProbs]
+    finish_reason: str
+    prompt_tokens: int
+    completion_tokens: int
+    cached_tokens: int
+    output_logprobs: OutputLogProbs
     all_hidden_states: _containers.RepeatedCompositeFieldContainer[HiddenStates]
-    def __init__(self, output_ids: _Optional[_Iterable[int]] = ..., output_text: _Optional[str] = ..., finish_reason: _Optional[_Union[GenerateComplete.FinishReason, str]] = ..., all_logprobs: _Optional[_Iterable[_Union[LogProbs, _Mapping]]] = ..., all_hidden_states: _Optional[_Iterable[_Union[HiddenStates, _Mapping]]] = ...) -> None: ...
+    matched_token_id: int
+    matched_stop_str: str
+    input_logprobs: InputLogProbs
+    index: int
+    def __init__(self, output_ids: _Optional[_Iterable[int]] = ..., finish_reason: _Optional[str] = ..., prompt_tokens: _Optional[int] = ..., completion_tokens: _Optional[int] = ..., cached_tokens: _Optional[int] = ..., output_logprobs: _Optional[_Union[OutputLogProbs, _Mapping]] = ..., all_hidden_states: _Optional[_Iterable[_Union[HiddenStates, _Mapping]]] = ..., matched_token_id: _Optional[int] = ..., matched_stop_str: _Optional[str] = ..., input_logprobs: _Optional[_Union[InputLogProbs, _Mapping]] = ..., index: _Optional[int] = ...) -> None: ...
 
 class GenerateError(_message.Message):
     __slots__ = ("message", "http_status_code", "details")
@@ -218,27 +215,39 @@ class GenerateError(_message.Message):
     details: str
     def __init__(self, message: _Optional[str] = ..., http_status_code: _Optional[str] = ..., details: _Optional[str] = ...) -> None: ...
 
-class LogProbs(_message.Message):
-    __slots__ = ("token_logprobs", "token_ids", "top_logprobs", "token_texts")
+class OutputLogProbs(_message.Message):
+    __slots__ = ("token_logprobs", "token_ids", "top_logprobs")
     TOKEN_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
     TOKEN_IDS_FIELD_NUMBER: _ClassVar[int]
     TOP_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
-    TOKEN_TEXTS_FIELD_NUMBER: _ClassVar[int]
     token_logprobs: _containers.RepeatedScalarFieldContainer[float]
     token_ids: _containers.RepeatedScalarFieldContainer[int]
     top_logprobs: _containers.RepeatedCompositeFieldContainer[TopLogProbs]
-    token_texts: _containers.RepeatedScalarFieldContainer[str]
-    def __init__(self, token_logprobs: _Optional[_Iterable[float]] = ..., token_ids: _Optional[_Iterable[int]] = ..., top_logprobs: _Optional[_Iterable[_Union[TopLogProbs, _Mapping]]] = ..., token_texts: _Optional[_Iterable[str]] = ...) -> None: ...
+    def __init__(self, token_logprobs: _Optional[_Iterable[float]] = ..., token_ids: _Optional[_Iterable[int]] = ..., top_logprobs: _Optional[_Iterable[_Union[TopLogProbs, _Mapping]]] = ...) -> None: ...
+
+class InputLogProbs(_message.Message):
+    __slots__ = ("token_logprobs", "token_ids", "top_logprobs")
+    TOKEN_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
+    TOKEN_IDS_FIELD_NUMBER: _ClassVar[int]
+    TOP_LOGPROBS_FIELD_NUMBER: _ClassVar[int]
+    token_logprobs: _containers.RepeatedCompositeFieldContainer[InputTokenLogProb]
+    token_ids: _containers.RepeatedScalarFieldContainer[int]
+    top_logprobs: _containers.RepeatedCompositeFieldContainer[TopLogProbs]
+    def __init__(self, token_logprobs: _Optional[_Iterable[_Union[InputTokenLogProb, _Mapping]]] = ..., token_ids: _Optional[_Iterable[int]] = ..., top_logprobs: _Optional[_Iterable[_Union[TopLogProbs, _Mapping]]] = ...) -> None: ...
+
+class InputTokenLogProb(_message.Message):
+    __slots__ = ("value",)
+    VALUE_FIELD_NUMBER: _ClassVar[int]
+    value: float
+    def __init__(self, value: _Optional[float] = ...) -> None: ...
 
 class TopLogProbs(_message.Message):
-    __slots__ = ("values", "token_ids", "token_texts")
+    __slots__ = ("values", "token_ids")
     VALUES_FIELD_NUMBER: _ClassVar[int]
     TOKEN_IDS_FIELD_NUMBER: _ClassVar[int]
-    TOKEN_TEXTS_FIELD_NUMBER: _ClassVar[int]
     values: _containers.RepeatedScalarFieldContainer[float]
     token_ids: _containers.RepeatedScalarFieldContainer[int]
-    token_texts: _containers.RepeatedScalarFieldContainer[str]
-    def __init__(self, values: _Optional[_Iterable[float]] = ..., token_ids: _Optional[_Iterable[int]] = ..., token_texts: _Optional[_Iterable[str]] = ...) -> None: ...
+    def __init__(self, values: _Optional[_Iterable[float]] = ..., token_ids: _Optional[_Iterable[int]] = ...) -> None: ...
 
 class HiddenStates(_message.Message):
     __slots__ = ("values", "layer", "position")
@@ -283,20 +292,18 @@ class EmbedResponse(_message.Message):
     def __init__(self, request_id: _Optional[str] = ..., complete: _Optional[_Union[EmbedComplete, _Mapping]] = ..., error: _Optional[_Union[EmbedError, _Mapping]] = ...) -> None: ...
 
 class EmbedComplete(_message.Message):
-    __slots__ = ("embedding", "prompt_tokens", "cached_tokens", "embedding_dim", "generation_time", "batch_embeddings")
+    __slots__ = ("embedding", "prompt_tokens", "cached_tokens", "embedding_dim", "batch_embeddings")
     EMBEDDING_FIELD_NUMBER: _ClassVar[int]
     PROMPT_TOKENS_FIELD_NUMBER: _ClassVar[int]
     CACHED_TOKENS_FIELD_NUMBER: _ClassVar[int]
     EMBEDDING_DIM_FIELD_NUMBER: _ClassVar[int]
-    GENERATION_TIME_FIELD_NUMBER: _ClassVar[int]
     BATCH_EMBEDDINGS_FIELD_NUMBER: _ClassVar[int]
     embedding: _containers.RepeatedScalarFieldContainer[float]
     prompt_tokens: int
     cached_tokens: int
     embedding_dim: int
-    generation_time: float
     batch_embeddings: _containers.RepeatedCompositeFieldContainer[Embedding]
-    def __init__(self, embedding: _Optional[_Iterable[float]] = ..., prompt_tokens: _Optional[int] = ..., cached_tokens: _Optional[int] = ..., embedding_dim: _Optional[int] = ..., generation_time: _Optional[float] = ..., batch_embeddings: _Optional[_Iterable[_Union[Embedding, _Mapping]]] = ...) -> None: ...
+    def __init__(self, embedding: _Optional[_Iterable[float]] = ..., prompt_tokens: _Optional[int] = ..., cached_tokens: _Optional[int] = ..., embedding_dim: _Optional[int] = ..., batch_embeddings: _Optional[_Iterable[_Union[Embedding, _Mapping]]] = ...) -> None: ...
 
 class Embedding(_message.Message):
     __slots__ = ("values", "index")

sglang/srt/grpc/sglang_scheduler_pb2_grpc.py

@@ -1,3 +1,6 @@
+# This file is auto-generated. Do not edit manually.
+# Regenerate with: python compile_proto.py
+
 # Generated by the gRPC Python protocol compiler plugin. DO NOT EDIT!
 """Client and server classes corresponding to protobuf-defined services."""
 import grpc

sglang/srt/layers/activation.py

@@ -224,12 +224,13 @@ class XIELU(CustomOp):
                 self._xielu_cuda_fn = self._xielu_cuda
             logger.warning_once(msg)
         except Exception as err:
-            logger.warning_once(
-                "CUDA-fused xIELU not available (%s) –"
-                " falling back to a Python version.\n"
-                "For CUDA xIELU (experimental), `pip install git+https://github.com/nickjbrowning/XIELU`",
-                str(err),
-            )
+            pass
+            # logger.warning_once(
+            #     "CUDA-fused xIELU not available (%s) –"
+            #     " falling back to a Python version.\n"
+            #     "For CUDA xIELU (experimental), `pip install git+https://github.com/nickjbrowning/XIELU`",
+            #     str(err),
+            # )
 
     def _xielu_python(self, x: torch.Tensor) -> torch.Tensor:
         alpha_p = nn.functional.softplus(self.alpha_p)

sglang/srt/layers/attention/aiter_backend.py

@@ -4,18 +4,13 @@ from __future__ import annotations
 end to end attention solution with aiter kernels
 """
 
-import math
-import os
 from dataclasses import dataclass
 from enum import Enum, auto
-from functools import partial
-from typing import TYPE_CHECKING, List, Optional, Union
+from typing import TYPE_CHECKING, Optional
 
 import torch
 import triton
-import triton.language as tl
 
-from sglang.global_config import global_config
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.layers.dp_attention import (
@@ -27,7 +22,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMo
 if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention
     from sglang.srt.model_executor.model_runner import ModelRunner
-    from sglang.srt.speculative.spec_info import SpecInfo
+    from sglang.srt.speculative.spec_info import SpecInput
 
 try:
     from aiter import (
@@ -374,7 +369,7 @@ class AiterAttnBackend(AttentionBackend):
         seq_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
     ):
         if forward_mode.is_decode_or_idle():
             qo_indptr = None
@@ -509,7 +504,7 @@ class AiterAttnBackend(AttentionBackend):
         seq_lens_sum: int,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
         seq_lens_cpu: Optional[torch.Tensor],
     ):
         if forward_mode.is_decode_or_idle():
@@ -619,7 +614,11 @@ class AiterAttnBackend(AttentionBackend):
             assert len(k.shape) == 3
             assert len(v.shape) == 3
 
-            if forward_batch.forward_mode.is_extend():
+            if (
+                forward_batch.forward_mode.is_extend()
+                and not forward_batch.forward_mode.is_target_verify()
+                and not forward_batch.forward_mode.is_draft_extend()
+            ):
                 if kv_indices.shape[0] == 0:
                     o = flash_attn_varlen_func(
                         q,
@@ -884,7 +883,7 @@ class AiterIndicesUpdaterPrefill:
         seq_lens_sum: int,
         prefix_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
     ):
         # Keep the signature for type checking. It will be assigned during runtime.
         raise NotImplementedError()
@@ -896,7 +895,7 @@ class AiterIndicesUpdaterPrefill:
         seq_lens_sum: int,
         prefix_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
     ):
 
         kv_start_idx = None
@@ -980,7 +979,7 @@ class AiterMlaIndicesUpdaterPrefill:
         extend_lens: torch.Tensor,
         max_q_len: int,
         max_kv_len: int,
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
     ):
         # Keep the signature for type checking. It will be assigned during runtime.
         raise NotImplementedError()
@@ -993,7 +992,7 @@ class AiterMlaIndicesUpdaterPrefill:
         extend_lens: torch.Tensor,
         max_q_len: int,
         max_kv_len: int,
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
     ):
         bs = len(req_pool_indices)
 
@@ -1050,7 +1049,7 @@ class AiterMultiStepDraftBackend:
         topk: int,
         speculative_num_steps: int,
     ):
-        from sglang.srt.speculative.eagle_utils import generate_draft_decode_kv_indices
+        from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
 
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps

sglang/srt/layers/attention/ascend_backend.py

@@ -5,14 +5,15 @@ from typing import TYPE_CHECKING, List, Optional
 
 import torch
 import torch_npu
-from torch.nn.functional import scaled_dot_product_attention
 
 from sglang.srt.configs.model_config import AttentionArch
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
+from sglang.srt.layers.attention.npu_ops.mla_preprocess import is_mla_preprocess_enabled
 from sglang.srt.layers.attention.torch_native_backend import TorchNativeAttnBackend
 from sglang.srt.layers.dp_attention import get_attention_tp_size
 from sglang.srt.layers.radix_attention import AttentionType
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
+from sglang.srt.speculative.spec_info import SpecInput
 from sglang.srt.utils import get_bool_env_var
 
 if TYPE_CHECKING:
@@ -35,6 +36,8 @@ class ForwardMetadata:
     seq_lens_cpu_int: Optional[torch.Tensor] = None
     seq_lens_cpu_list: Optional[List[int]] = None
     seq_lens_list_cumsum: Optional[List[int]] = None
+    seq_lens: Optional[torch.Tensor] = None
+    actual_seq_lengths_q: Optional[torch.Tensor] = None
 
 
 class AscendAttnBackend(AttentionBackend):
@@ -66,6 +69,9 @@ class AscendAttnBackend(AttentionBackend):
         if self.use_mla:
             self.kv_lora_rank = model_runner.model_config.kv_lora_rank
             self.qk_rope_head_dim = model_runner.model_config.qk_rope_head_dim
+            self.q_head_dim = (
+                self.qk_rope_head_dim + model_runner.model_config.qk_nope_head_dim
+            )
         self.native_attn = TorchNativeAttnBackend(model_runner)
         self.graph_metadata = {}
         self.max_context_len = model_runner.model_config.context_len
@@ -101,10 +107,6 @@ class AscendAttnBackend(AttentionBackend):
         self.forward_metadata.seq_lens_cpu_int = forward_batch.seq_lens_cpu.int()
 
         seq_lens_list_cumsum = np.cumsum(forward_batch.extend_seq_lens_cpu)
-        if forward_batch.is_extend_in_batch:
-            seq_lens_list_cumsum[-1] = (
-                (seq_lens_list_cumsum[-1] - 1) // tp_size + 1
-            ) * tp_size
         self.forward_metadata.seq_lens_list_cumsum = seq_lens_list_cumsum
 
         self.graph_mode = False
@@ -126,12 +128,16 @@ class AscendAttnBackend(AttentionBackend):
         seq_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]],
+        spec_info: Optional[SpecInput],
     ):
         metadata = ForwardMetadata()
 
         metadata.block_tables = self.graph_metadata["block_tables"][:bs, :]
         metadata.seq_lens_cpu_list = seq_lens.cpu().int().tolist()
+        metadata.seq_lens = seq_lens
+        metadata.actual_seq_lengths_q = torch.tensor(
+            [1 + i * 1 for i in range(bs)], dtype=torch.int32, device=seq_lens.device
+        )
 
         self.graph_metadata[bs] = metadata
         self.forward_metadata = metadata
@@ -146,7 +152,7 @@ class AscendAttnBackend(AttentionBackend):
         seq_lens_sum: int,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]],
+        spec_info: Optional[SpecInput],
         seq_lens_cpu: Optional[torch.Tensor],
     ):
         metadata = self.graph_metadata[bs]
@@ -160,6 +166,8 @@ class AscendAttnBackend(AttentionBackend):
         metadata.block_tables[:bs, max_seq_pages:].fill_(0)
         metadata.block_tables[bs:, :].fill_(0)
 
+        metadata.seq_lens[:bs].copy_(seq_lens[:bs])
+
         self.forward_metadata = metadata
 
         self.graph_mode = True
@@ -167,6 +175,64 @@ class AscendAttnBackend(AttentionBackend):
     def get_cuda_graph_seq_len_fill_value(self):
         return 0
 
+    def forward_sparse(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache: bool = True,
+        # For multi_head latent attention
+        q_rope: Optional[torch.Tensor] = None,
+        k_rope: Optional[torch.Tensor] = None,
+        topk_indices: torch.Tensor = None,
+    ):
+
+        is_prefill = forward_batch.forward_mode.is_extend()
+
+        if save_kv_cache:
+            k = k.view(-1, layer.tp_k_head_num, self.kv_lora_rank)
+            k_rope = k_rope.view(-1, layer.tp_k_head_num, self.qk_rope_head_dim)
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, k_rope
+            )
+        q_nope, q_pe = q, q_rope
+        k_nope, k_pe = forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id)
+        block_table = self.forward_metadata.block_tables
+        if is_prefill:
+            actual_seq_qlen = torch.cumsum(forward_batch.seq_lens, dim=0)
+        else:
+            if self.forward_metadata.actual_seq_lengths_q is None:
+                actual_seq_qlen = (
+                    torch.arange(1, q.shape[0] + 1).to(q.device).to(torch.int32)
+                )
+            else:
+                actual_seq_qlen = self.forward_metadata.actual_seq_lengths_q
+        if self.forward_metadata.seq_lens_cpu_int is None:
+            actual_seq_lengths_kv = self.forward_metadata.seq_lens
+        else:
+            actual_seq_lengths_kv = self.forward_metadata.seq_lens_cpu_int
+
+        attn_out = torch.ops.custom.npu_sparse_flash_attention(
+            query=q_nope,
+            key=k_nope,
+            value=k_nope,
+            query_rope=q_pe,
+            key_rope=k_pe,
+            sparse_indices=topk_indices,
+            scale_value=layer.scaling,
+            actual_seq_lengths_query=actual_seq_qlen.to(torch.int32),
+            actual_seq_lengths_kv=actual_seq_lengths_kv.to(q.device),
+            block_table=block_table,
+            sparse_block_size=1,
+            layout_query="TND",
+            layout_kv="PA_BSND",
+            sparse_mode=3,
+        )
+
+        return attn_out
+
     def forward_extend(
         self,
         q,
@@ -175,7 +241,23 @@ class AscendAttnBackend(AttentionBackend):
         layer: RadixAttention,
         forward_batch: ForwardBatch,
         save_kv_cache: bool = True,
+        # For multi_head latent attention
+        q_rope: Optional[torch.Tensor] = None,
+        k_rope: Optional[torch.Tensor] = None,
+        topk_indices: Optional[torch.Tensor] = None,
     ):
+        if topk_indices is not None:
+            return self.forward_sparse(
+                q,
+                k,
+                v,
+                layer,
+                forward_batch,
+                save_kv_cache,
+                q_rope,
+                k_rope,
+                topk_indices,
+            )
         if not self.use_mla:
             if save_kv_cache:
                 forward_batch.token_to_kv_pool.set_kv_buffer(
@@ -401,7 +483,7 @@ class AscendAttnBackend(AttentionBackend):
                 antiquant_scale=None,
                 sparse_mode=0,
             )
-            output = torch.zeros_like(q_nope, dtype=q.dtype, device=q.device)
+            output = torch.empty_like(q_nope, dtype=q.dtype, device=q.device)
             softmax_lse = torch.empty(1, dtype=q.dtype, device=q.device)
 
             torch_npu.npu_fused_infer_attention_score.out(
@@ -436,7 +518,24 @@ class AscendAttnBackend(AttentionBackend):
         # For multi-head latent attention
         q_rope: Optional[torch.Tensor] = None,
         k_rope: Optional[torch.Tensor] = None,
+        topk_indices: Optional[torch.Tensor] = None,
     ):
+        if is_mla_preprocess_enabled():
+            # MLAPO does saving kv_cache
+            save_kv_cache = False
+        if topk_indices is not None:
+            return self.forward_sparse(
+                q,
+                k,
+                v,
+                layer,
+                forward_batch,
+                save_kv_cache,
+                q_rope,
+                k_rope,
+                topk_indices,
+            )
+
         if self.graph_mode:
             return self.forward_decode_graph(
                 q,