sglang 0.5.3rc2__py3-none-any.whl → 0.5.4.post1__py3-none-any.whl

sglang/srt/models/dots_ocr.py

@@ -6,7 +6,6 @@ from typing import Iterable, List, Optional, Tuple
 
 import torch
 import torch.nn as nn
-from transformers.activations import ACT2FN
 
 from sglang.srt.configs import DotsOCRConfig
 from sglang.srt.layers.logits_processor import LogitsProcessor
@@ -22,7 +21,6 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.dots_vlm_vit import DotsVisionTransformer
 from sglang.srt.models.qwen2 import Qwen2ForCausalLM
 from sglang.srt.utils import add_prefix
-from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
 

sglang/srt/models/dots_vlm.py

@@ -23,7 +23,6 @@ import torch
 from torch import nn
 
 from sglang.srt.configs.dots_vlm import DotsVLMConfig
-from sglang.srt.distributed import parallel_state
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.managers.mm_utils import (
     MultiModalityDataPaddingPatternMultimodalTokens,

sglang/srt/models/dots_vlm_vit.py

@@ -323,7 +323,7 @@ class DotsVisionTransformer(PreTrainedModel):
             dim=0,
             dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
         )
-        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+        cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
 
         for blk in self.blocks:
             hidden_states = blk(

sglang/srt/models/falcon_h1.py

@@ -1,4 +1,3 @@
-import enum
 import logging
 from typing import Any, Iterable, List, Optional, Set, Tuple
 
@@ -8,6 +7,10 @@ from torch import nn
 from sglang.srt.configs.falcon_h1 import FalconH1Config
 from sglang.srt.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.attention.hybrid_linear_attn_backend import (
+    HybridLinearAttnBackend,
+    Mamba2AttnBackend,
+)
 from sglang.srt.layers.attention.mamba.mamba import MambaMixer2
 from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
 from sglang.srt.layers.dp_attention import (
@@ -29,9 +32,9 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, is_cuda, make_layers
 
 logger = logging.getLogger(__name__)
@@ -184,18 +187,12 @@ class FalconH1HybridAttentionDecoderLayer(nn.Module):
         )
 
         self.mamba = MambaMixer2(
+            cache_params=config.mamba2_cache_params,
             hidden_size=config.hidden_size,
-            ssm_state_size=config.mamba_d_state,
-            conv_kernel_size=config.mamba_d_conv,
-            intermediate_size=self.d_ssm,
             use_conv_bias=config.mamba_conv_bias,
             use_bias=config.mamba_proj_bias,
             n_groups=config.mamba_n_groups,
-            num_heads=config.mamba_n_heads,
-            layer_id=layer_id,
-            head_dim=config.mamba_d_head,
             rms_norm_eps=config.rms_norm_eps,
-            chunk_size=config.mamba_chunk_size,
             activation=config.hidden_act,
             use_rms_norm=config.mamba_rms_norm,
             prefix=f"{prefix}.mixer",
@@ -339,12 +336,16 @@ class FalconH1HybridAttentionDecoderLayer(nn.Module):
             )
             attention_hidden_states = attention_hidden_states * self.attn_out_multiplier
 
+            attn_backend = forward_batch.attn_backend
+            assert isinstance(attn_backend, HybridLinearAttnBackend)
+            assert isinstance(attn_backend.linear_attn_backend, Mamba2AttnBackend)
             # Mamba block
             mamba_hidden_states = torch.empty_like(hidden_states)
-            self.mamba(
+            attn_backend.linear_attn_backend.forward(
+                self.mamba,
                 hidden_states * self.ssm_in_multiplier,
                 mamba_hidden_states,
-                forward_batch=forward_batch,
+                layer_id=self.layer_id,
                 mup_vector=self.mup_vector,
             )
             mamba_hidden_states = mamba_hidden_states * self.ssm_out_multiplier
@@ -448,13 +449,6 @@ class FalconH1Model(nn.Module):
         return hidden_states
 
 
-class HybridLayerType(enum.Enum):
-    full_attention = "attention"
-    swa_attention = "swa_attention"
-    linear_attention = "linear_attention"
-    mamba2 = "mamba"
-
-
 class FalconH1ForCausalLM(nn.Module):
     fall_back_to_pt_during_load = False
 
@@ -481,7 +475,7 @@ class FalconH1ForCausalLM(nn.Module):
                 quant_config=quant_config,
                 org_num_embeddings=config.vocab_size,
                 prefix=add_prefix("lm_head", prefix),
-                use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
+                use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
             )
         self.lm_head = self.lm_head.float()
         self.lm_head_multiplier = config.lm_head_multiplier

sglang/srt/models/gemma3_mm.py

@@ -16,6 +16,7 @@
 # https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/gemma3_mm.py
 
 import logging
+import re
 from functools import lru_cache
 from typing import Dict, Iterable, List, Optional, Set, Tuple, TypedDict
 
@@ -154,6 +155,10 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
     embedding_modules = {}
     embedding_padding_modules = []
     supports_lora = True
+    # Pattern to match language model layers only (skip vision_tower and multi_modal_projector)
+    lora_pattern = re.compile(
+        r"^language_model\.model\.layers\.(\d+)\.(?:self_attn|mlp)\.(?:qkv_proj|o_proj|down_proj|gate_up_proj)"
+    )
 
     def __init__(
         self,
@@ -165,6 +170,13 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
         self.config = config
         self.quant_config = quant_config
 
+        # For LoRA compatibility: expose text_config attributes at top level
+        # This allows LoRA code to work without special multimodal handling
+        if not hasattr(config, "num_hidden_layers"):
+            config.num_hidden_layers = config.text_config.num_hidden_layers
+        if not hasattr(config, "hidden_size"):
+            config.hidden_size = config.text_config.hidden_size
+
         self.vision_tower = SiglipVisionModel(
             config=config.vision_config,
             quant_config=quant_config,
@@ -380,6 +392,10 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
 
         return hs
 
+    def should_apply_lora(self, module_name: str) -> bool:
+        """Skip vision tower and multi_modal_projector for LoRA."""
+        return bool(self.lora_pattern.match(module_name))
+
     def tie_weights(self):
         return self.language_model.tie_weights()
 

sglang/srt/models/gemma3n_mm.py

@@ -14,8 +14,7 @@ from transformers import (
 )
 from transformers.models.auto.modeling_auto import AutoModel
 
-from sglang.srt.layers.layernorm import RMSNorm
-from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
+from sglang.srt.layers.linear import RowParallelLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding