sglang 0.4.8__py3-none-any.whl → 0.4.9__py3-none-any.whl

sglang/srt/models/qwen2_audio.py

@@ -0,0 +1,200 @@
+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/1d45d90e5d1552eccb6d8cc9b7bba283ccefb808/src/transformers/models/qwen2_audio/modeling_qwen2_audio.py
+# Copyright 2024 The Qwen team.
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only Qwen2-Audio model compatible with HuggingFace weights."""
+import logging
+import math
+from functools import lru_cache, partial
+from typing import Any, Iterable, List, Optional, Tuple, Type, TypedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from transformers import AutoTokenizer, Qwen2AudioEncoderConfig, Qwen2Config
+from transformers.activations import ACT2FN
+from transformers.models.qwen2_audio.configuration_qwen2_audio import Qwen2AudioConfig
+from transformers.models.qwen2_audio.modeling_qwen2_audio import (
+    Qwen2AudioEncoder,
+    Qwen2AudioMultiModalProjector,
+)
+
+from sglang.srt.hf_transformers_utils import get_processor
+from sglang.srt.layers.activation import QuickGELU
+from sglang.srt.layers.attention.vision import VisionAttention
+from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.pooler import Pooler, PoolingType
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.utils import get_layer_id
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+from sglang.srt.managers.mm_utils import (
+    MultiModalityDataPaddingPatternMultimodalTokens,
+    general_mm_embed_routine,
+)
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.models.qwen2 import Qwen2ForCausalLM
+from sglang.srt.utils import add_prefix
+
+logger = logging.getLogger(__name__)
+
+
+class Qwen2AudioForConditionalGeneration(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
+    def __init__(
+        self,
+        config: Qwen2AudioConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+
+        if getattr(self.config, "audio_config", None) is None:
+            self.config.audio_config = Qwen2AudioEncoderConfig(
+                self.config._name_or_path
+            )
+
+        if getattr(self.config, "text_config", None) is None:
+            self.config.text_config = Qwen2Config(self.config._name_or_path)
+
+        self.audio_tower = Qwen2AudioEncoder(
+            config.audio_config,
+        )
+        self.multi_modal_projector = Qwen2AudioMultiModalProjector(config)
+        self.language_model = Qwen2ForCausalLM(
+            config.text_config, quant_config, prefix=add_prefix("model", prefix)
+        )
+
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        # Get all special token IDs for audio
+        audio_token_id: int = getattr(
+            mm_inputs, "audio_token_id", mm_inputs.im_token_id
+        )
+
+        pattern = MultiModalityDataPaddingPatternMultimodalTokens([audio_token_id])
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
+
+    def get_audio_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        # Extract audio features from input items
+        input_features = torch.cat([item.audio_features for item in items], dim=0).type(
+            self.audio_tower.dtype
+        )
+
+        audio_embeds = self.audio_tower(input_features).last_hidden_state
+        audio_embeds = self.multi_modal_projector(audio_embeds)
+
+        audio_feature_lens = torch.cat([item.audio_feature_lens for item in items])
+        new_embeds = []
+        for i, d in zip(audio_feature_lens, audio_embeds):
+            new_embeds.append(d[: i.item()])
+
+        return torch.cat(new_embeds, dim=0)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        **kwargs: Any,
+    ) -> torch.Tensor:
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            language_model=self.language_model,
+            audio_data_embedding_func=self.get_audio_feature,
+            positions=positions,
+        )
+
+        return hidden_states
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+
+            if self.config.text_config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name or "audio_tower" in name:
+                    continue
+                name_tmp = name.replace(weight_name, param_name)
+
+                # Skip loading extra bias for GPTQ models.
+                if name_tmp.endswith(".bias") and name_tmp not in params_dict:
+                    continue
+                param = params_dict[name_tmp]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                try:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    param = params_dict[name]
+                except KeyError:
+                    print(params_dict.keys())
+                    raise
+
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+
+EntryClass = Qwen2AudioForConditionalGeneration

sglang/srt/models/qwen2_moe.py

@@ -31,6 +31,11 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
+from sglang.srt.eplb.expert_distribution import (
+    ExpertDistributionRecorder,
+    get_global_expert_distribution_recorder,
+)
+from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.communicator import (
     LayerCommunicator,
@@ -64,11 +69,6 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.expert_distribution import (
-    ExpertDistributionRecorder,
-    get_global_expert_distribution_recorder,
-)
-from sglang.srt.managers.expert_location import ModelConfigForExpertLocation
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
@@ -143,6 +143,15 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
+            # Additional args for FusedMoE
+            **(
+                dict(
+                    enable_flashinfer_moe=True,
+                    enable_ep_moe=global_server_args_dict["enable_ep_moe"],
+                )
+                if global_server_args_dict["enable_flashinfer_moe"]
+                else {}
+            ),
         )
 
         self.gate = ReplicatedLinear(
@@ -291,6 +300,7 @@ class Qwen2MoeDecoderLayer(nn.Module):
         layer_id: int,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -393,6 +403,7 @@ class Qwen2MoeModel(nn.Module):
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
         decoder_layer_type: type[nn.Module] = Qwen2MoeDecoderLayer,
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.padding_idx = config.pad_token_id
@@ -418,6 +429,7 @@ class Qwen2MoeModel(nn.Module):
                 config=config,
                 quant_config=quant_config,
                 prefix=prefix,
+                alt_stream=alt_stream,
             ),
             pp_rank=self.pp_group.rank_in_group,
             pp_size=self.pp_group.world_size,
@@ -428,6 +440,9 @@ class Qwen2MoeModel(nn.Module):
         else:
             self.norm = PPMissingLayer(return_tuple=True)
 
+        # For EAGLE3 support
+        self.layers_to_capture = []
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -447,6 +462,7 @@ class Qwen2MoeModel(nn.Module):
             hidden_states = pp_proxy_tensors["hidden_states"]
             residual = pp_proxy_tensors["residual"]
 
+        aux_hidden_states = []
         if forward_batch.can_run_tbo:
             hidden_states, residual = model_forward_maybe_tbo(
                 layers=self.layers,
@@ -459,6 +475,12 @@ class Qwen2MoeModel(nn.Module):
             )
         else:
             for i in range(self.start_layer, self.end_layer):
+                if i in self.layers_to_capture:
+                    aux_hidden_states.append(
+                        hidden_states + residual
+                        if residual is not None
+                        else hidden_states
+                    )
                 with get_global_expert_distribution_recorder().with_current_layer(i):
                     layer = self.layers[i]
                     hidden_states, residual = layer(
@@ -477,7 +499,11 @@ class Qwen2MoeModel(nn.Module):
                     hidden_states = self.norm(hidden_states)
                 else:
                     hidden_states, _ = self.norm(hidden_states, residual)
-        return hidden_states
+
+        if len(aux_hidden_states) == 0:
+            return hidden_states
+
+        return hidden_states, aux_hidden_states
 
 
 class Qwen2MoeForCausalLM(nn.Module):

sglang/srt/models/qwen2_vl.py

@@ -479,10 +479,7 @@ class Qwen2VLForConditionalGeneration(nn.Module):
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
     def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
-        # Get all special token IDs
-        im_token_id: int = mm_inputs.im_token_id
-
-        pattern = MultiModalityDataPaddingPatternMultimodalTokens([im_token_id])
+        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
         return pattern.pad_input_tokens(input_ids, mm_inputs)
 
     def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:

sglang/srt/models/qwen3.py

@@ -2,7 +2,7 @@
 
 import logging
 from functools import partial
-from typing import Any, Dict, Iterable, Optional, Tuple
+from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
 from torch import nn
@@ -11,9 +11,9 @@ from sglang.srt.distributed import (
     get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
-    split_tensor_along_last_dim,
-    tensor_model_parallel_all_gather,
 )
+from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
+from sglang.srt.layers.dp_attention import get_attention_tp_rank, get_attention_tp_size
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
@@ -23,15 +23,17 @@ from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
 from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2 import Qwen2MLP as Qwen3MLP
 from sglang.srt.models.qwen2 import Qwen2Model
-from sglang.srt.utils import add_prefix
+from sglang.srt.utils import add_prefix, is_cuda
 
 Qwen3Config = None
 
 logger = logging.getLogger(__name__)
+_is_cuda = is_cuda()
 
 
 class Qwen3Attention(nn.Module):
@@ -49,23 +51,27 @@ class Qwen3Attention(nn.Module):
         rms_norm_eps: float = None,
         attention_bias: bool = False,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
         self.tp_size = get_tensor_model_parallel_world_size()
         self.total_num_heads = num_heads
-        assert self.total_num_heads % self.tp_size == 0
-        self.num_heads = self.total_num_heads // self.tp_size
+        attn_tp_rank = get_attention_tp_rank()
+        attn_tp_size = get_attention_tp_size()
+
+        assert self.total_num_heads % attn_tp_size == 0
+        self.num_heads = self.total_num_heads // attn_tp_size
         self.total_num_kv_heads = num_kv_heads
-        if self.total_num_kv_heads >= self.tp_size:
+        if self.total_num_kv_heads >= attn_tp_size:
             # Number of KV heads is greater than TP size, so we partition
             # the KV heads across multiple tensor parallel GPUs.
-            assert self.total_num_kv_heads % self.tp_size == 0
+            assert self.total_num_kv_heads % attn_tp_size == 0
         else:
             # Number of KV heads is less than TP size, so we replicate
             # the KV heads across multiple tensor parallel GPUs.
-            assert self.tp_size % self.total_num_kv_heads == 0
-        self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size)
+            assert attn_tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size)
         self.head_dim = head_dim or hidden_size // self.total_num_heads
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
@@ -84,6 +90,8 @@ class Qwen3Attention(nn.Module):
             self.total_num_kv_heads,
             bias=attention_bias,
             quant_config=quant_config,
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
             prefix=add_prefix("qkv_proj", prefix),
         )
         self.o_proj = RowParallelLinear(
@@ -91,6 +99,9 @@ class Qwen3Attention(nn.Module):
             hidden_size,
             bias=attention_bias,
             quant_config=quant_config,
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
+            reduce_results=False,
             prefix=add_prefix("o_proj", prefix),
         )
 
@@ -109,15 +120,27 @@ class Qwen3Attention(nn.Module):
             layer_id=layer_id,
             prefix=add_prefix("attn", prefix),
         )
+        self.alt_stream = alt_stream
 
     def _apply_qk_norm(
         self, q: torch.Tensor, k: torch.Tensor
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        q_by_head = q.reshape(-1, self.head_dim)
-        q_by_head = self.q_norm(q_by_head)
+        # overlap qk norm
+        if self.alt_stream is not None and get_is_capture_mode():
+            current_stream = torch.cuda.current_stream()
+            self.alt_stream.wait_stream(current_stream)
+            q_by_head = q.reshape(-1, self.head_dim)
+            q_by_head = self.q_norm(q_by_head)
+            with torch.cuda.stream(self.alt_stream):
+                k_by_head = k.reshape(-1, self.head_dim)
+                k_by_head = self.k_norm(k_by_head)
+            current_stream.wait_stream(self.alt_stream)
+        else:
+            q_by_head = q.reshape(-1, self.head_dim)
+            q_by_head = self.q_norm(q_by_head)
+            k_by_head = k.reshape(-1, self.head_dim)
+            k_by_head = self.k_norm(k_by_head)
         q = q_by_head.view(q.shape)
-        k_by_head = k.reshape(-1, self.head_dim)
-        k_by_head = self.k_norm(k_by_head)
         k = k_by_head.view(k.shape)
         return q, k
 
@@ -143,6 +166,7 @@ class Qwen3DecoderLayer(nn.Module):
         layer_id: int = 0,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -163,6 +187,7 @@ class Qwen3DecoderLayer(nn.Module):
             rms_norm_eps=config.rms_norm_eps,
             attention_bias=config.attention_bias,
             prefix=add_prefix("self_attn", prefix),
+            alt_stream=alt_stream,
         )
         self.mlp = Qwen3MLP(
             hidden_size=self.hidden_size,
@@ -176,6 +201,18 @@ class Qwen3DecoderLayer(nn.Module):
             config.hidden_size, eps=config.rms_norm_eps
         )
 
+        self.layer_scatter_modes = LayerScatterModes.init_new(
+            layer_id=layer_id,
+            num_layers=config.num_hidden_layers,
+            is_layer_sparse=False,
+            is_previous_layer_sparse=False,
+        )
+        self.layer_communicator = LayerCommunicator(
+            layer_scatter_modes=self.layer_scatter_modes,
+            input_layernorm=self.input_layernorm,
+            post_attention_layernorm=self.post_attention_layernorm,
+        )
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -184,20 +221,24 @@ class Qwen3DecoderLayer(nn.Module):
         residual: Optional[torch.Tensor],
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         # Self Attention
-        if residual is None:
-            residual = hidden_states
-            hidden_states = self.input_layernorm(hidden_states)
-        else:
-            hidden_states, residual = self.input_layernorm(hidden_states, residual)
-        hidden_states = self.self_attn(
-            positions=positions,
-            hidden_states=hidden_states,
-            forward_batch=forward_batch,
+        hidden_states, residual = self.layer_communicator.prepare_attn(
+            hidden_states, residual, forward_batch
         )
+        if hidden_states.shape[0] != 0:
+            hidden_states = self.self_attn(
+                positions=positions,
+                hidden_states=hidden_states,
+                forward_batch=forward_batch,
+            )
 
         # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states, residual = self.layer_communicator.prepare_mlp(
+            hidden_states, residual, forward_batch
+        )
         hidden_states = self.mlp(hidden_states)
+        hidden_states, residual = self.layer_communicator.postprocess_layer(
+            hidden_states, residual, forward_batch
+        )
         return hidden_states, residual
 
 
@@ -208,11 +249,13 @@ class Qwen3Model(Qwen2Model):
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
+        alt_stream = torch.cuda.Stream() if _is_cuda else None
         super().__init__(
             config=config,
             quant_config=quant_config,
             prefix=prefix,
             decoder_layer_type=Qwen3DecoderLayer,
+            alt_stream=alt_stream,
         )
 
 
@@ -282,6 +325,9 @@ class Qwen3ForCausalLM(nn.Module):
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
+        # For EAGLE3 support
+        self.capture_aux_hidden_states = False
+
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.model.get_input_embeddings(input_ids)
 
@@ -303,10 +349,18 @@ class Qwen3ForCausalLM(nn.Module):
             pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
         if self.pp_group.is_last_rank:
             if not get_embedding:
                 return self.logits_processor(
-                    input_ids, hidden_states, self.lm_head, forward_batch
+                    input_ids,
+                    hidden_states,
+                    self.lm_head,
+                    forward_batch,
+                    aux_hidden_states,
                 )
             else:
                 return self.pooler(hidden_states, forward_batch)
@@ -404,5 +458,20 @@ class Qwen3ForCausalLM(nn.Module):
     def load_kv_cache_scales(self, quantization_param_path: str) -> None:
         self.model.load_kv_cache_scales(quantization_param_path)
 
+    def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
+        if not self.pp_group.is_last_rank:
+            return
+
+        self.capture_aux_hidden_states = True
+        if layer_ids is None:
+            num_layers = self.config.num_hidden_layers
+            self.model.layers_to_capture = [
+                2,
+                num_layers // 2,
+                num_layers - 3,
+            ]  # Specific layers for EAGLE3 support
+        else:
+            self.model.layers_to_capture = [val + 1 for val in layer_ids]
+
 
 EntryClass = Qwen3ForCausalLM