sglang 0.4.6.post4__py3-none-any.whl → 0.4.6.post5__py3-none-any.whl

sglang/srt/models/mllama.py

@@ -836,7 +836,6 @@ class MllamaForConditionalGeneration(nn.Module):
             prefix="multi_modal_projector",
         )
         self.logits_processor = LogitsProcessor(config.text_config)
-        self.capture_mode = False
 
     def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
         pixel_values = torch.cat(
@@ -865,7 +864,6 @@ class MllamaForConditionalGeneration(nn.Module):
                 pixel_values = torch.cat(
                     [item.pixel_values for item in mm_input.mm_items], dim=0
                 )
-                # max_num_images = max(max_num_images, sum(1 if item.is_image() else 0 for item in mm_input.items))
                 max_num_images = max(max_num_images, pixel_values.shape[1])
 
                 max_num_tiles = max(max_num_tiles, pixel_values.shape[2])
@@ -970,6 +968,8 @@ class MllamaForConditionalGeneration(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
     ) -> Union[Tuple, CausalLMOutputWithPast]:
+        from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
+
         batched_images, batched_ar_ids, batched_ar_mask, encoder_lens_need = (
             self._batch_image_inputs(forward_batch)
         )
@@ -978,7 +978,7 @@ class MllamaForConditionalGeneration(nn.Module):
         cross_attention_mask = None
         cross_attention_states = None
 
-        if self.capture_mode:
+        if get_is_capture_mode():
             # NOTE: when doing cuda graph capture, we do not want to skip cross attention
             # Make is a constant value to avoid cuda graph capture issue
             skip_cross_attention = False

sglang/srt/models/qwen2.py

@@ -15,12 +15,14 @@
 # Adapted from llama2.py
 # Modify details for the adaptation of Qwen2 model.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
-from typing import Any, Dict, Iterable, Optional, Tuple
+import logging
+from typing import Any, Dict, Iterable, Optional, Tuple, Union
 
 import torch
 from torch import nn
 
 from sglang.srt.distributed import (
+    get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
 )
@@ -36,11 +38,12 @@ from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 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,
     VocabParallelEmbedding,
 )
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import (
     default_weight_loader,
     kv_cache_scales_loader,
@@ -50,6 +53,9 @@ from sglang.srt.utils import add_prefix, make_layers
 Qwen2Config = None
 
 
+logger = logging.getLogger(__name__)
+
+
 class Qwen2MLP(nn.Module):
     def __init__(
         self,
@@ -245,15 +251,21 @@ class Qwen2Model(nn.Module):
         self.config = config
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size
-        self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size,
-            config.hidden_size,
-            quant_config=quant_config,
-            prefix=add_prefix("embed_tokens", prefix),
-        )
+        self.pp_group = get_pp_group()
+
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
         # Use the provided decoder layer type or default to Qwen2DecoderLayer
         decoder_layer_type = decoder_layer_type or Qwen2DecoderLayer
-        self.layers = make_layers(
+        self.layers, self.start_layer, self.end_layer = make_layers(
             config.num_hidden_layers,
             lambda idx, prefix: decoder_layer_type(
                 layer_id=idx,
@@ -261,9 +273,14 @@ class Qwen2Model(nn.Module):
                 quant_config=quant_config,
                 prefix=prefix,
             ),
+            pp_rank=self.pp_group.rank_in_group,
+            pp_size=self.pp_group.world_size,
             prefix=add_prefix("layers", prefix),
         )
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        if self.pp_group.is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
 
     def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor:
         if hasattr(self.config, "scale_emb"):
@@ -280,13 +297,20 @@ class Qwen2Model(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
-    ) -> torch.Tensor:
-        if input_embeds is None:
-            hidden_states = self.embed_tokens(input_ids)
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[torch.Tensor, PPProxyTensors]:
+        if self.pp_group.is_first_rank:
+            if input_embeds is None:
+                hidden_states = self.embed_tokens(input_ids)
+            else:
+                hidden_states = input_embeds
+            residual = None
         else:
-            hidden_states = input_embeds
-        residual = None
-        for i in range(len(self.layers)):
+            assert pp_proxy_tensors is not None
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
             layer = self.layers[i]
             hidden_states, residual = layer(
                 positions,
@@ -294,7 +318,15 @@ class Qwen2Model(nn.Module):
                 forward_batch,
                 residual,
             )
-        hidden_states, _ = self.norm(hidden_states, residual)
+        if not self.pp_group.is_last_rank:
+            return PPProxyTensors(
+                {
+                    "hidden_states": hidden_states,
+                    "residual": residual,
+                }
+            )
+        else:
+            hidden_states, _ = self.norm(hidden_states, residual)
         return hidden_states
 
     # If this function is called, it should always initialize KV cache scale
@@ -348,6 +380,7 @@ class Qwen2ForCausalLM(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        self.pp_group = get_pp_group()
         self.config = config
         self.quant_config = quant_config
         self.model = Qwen2Model(
@@ -379,14 +412,33 @@ class Qwen2ForCausalLM(nn.Module):
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
         get_embedding: bool = False,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
-        if not get_embedding:
-            return self.logits_processor(
-                input_ids, hidden_states, self.lm_head, forward_batch
-            )
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
+        )
+
+        if self.pp_group.is_last_rank:
+            if not get_embedding:
+                return self.logits_processor(
+                    input_ids, hidden_states, self.lm_head, forward_batch
+                )
+            else:
+                return self.pooler(hidden_states, forward_batch)
         else:
-            return self.pooler(hidden_states, forward_batch)
+            return hidden_states
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
@@ -400,6 +452,17 @@ class Qwen2ForCausalLM(nn.Module):
 
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
+
             if "rotary_emb.inv_freq" in name or "projector" in name:
                 continue
             if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
@@ -426,9 +489,15 @@ class Qwen2ForCausalLM(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     continue
-                param = params_dict[name]
-                weight_loader = getattr(param, "weight_loader", default_weight_loader)
-                weight_loader(param, loaded_weight)
+
+                if name in params_dict.keys():
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+                else:
+                    logger.warning(f"Parameter {name} not found in params_dict")
 
     def get_embed_and_head(self):
         return self.model.embed_tokens.weight, self.lm_head.weight

sglang/srt/models/qwen2_5_vl.py

@@ -146,6 +146,8 @@ class Qwen2_5_VisionBlock(nn.Module):
             num_heads=num_heads,
             projection_size=dim,
             use_qkv_parallel=True,
+            rotary_embed="normal",
+            proj_bias=True,
             qkv_backend=qkv_backend,
             softmax_in_single_precision=softmax_in_single_precision,
             flatten_batch=flatten_batch,
@@ -497,6 +499,12 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
         return pattern.pad_input_tokens(input_ids, mm_inputs)
 
     def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        if any(item.precomputed_features is not None for item in items):
+            if not all(item.precomputed_features is not None for item in items):
+                raise NotImplementedError(
+                    "MM inputs where only some items are precomputed."
+                )
+            return torch.concat([item.precomputed_features for item in items])
         # in qwen-vl, last dim is the same
         pixel_values = torch.cat([item.pixel_values for item in items], dim=0).type(
             self.visual.dtype