sglang 0.4.5__py3-none-any.whl → 0.4.5.post1__py3-none-any.whl

sglang/srt/models/llama4.py

@@ -27,6 +27,13 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
+from sglang.srt.layers.dp_attention import (
+    dp_gather_partial,
+    dp_scatter,
+    get_attention_dp_size,
+    get_attention_tp_rank,
+    get_attention_tp_size,
+)
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
     QKVParallelLinear,
@@ -38,9 +45,10 @@ 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.vocab_parallel_embedding import 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.models.llama import LlamaForCausalLM, LlamaMLP
-from sglang.srt.utils import add_prefix, get_compiler_backend, make_layers
+from sglang.srt.utils import add_prefix, fast_topk, get_compiler_backend, make_layers
 
 logger = logging.getLogger(__name__)
 
@@ -55,7 +63,7 @@ class Llama4MoE(nn.Module):
         topk: int,
         renormalize: bool,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        router_scores_aK, router_indices_aK = torch.topk(gating_output, topk, dim=-1)
+        router_scores_aK, router_indices_aK = fast_topk(gating_output, topk, dim=-1)
         router_scores_aK = torch.sigmoid(router_scores_aK.float()).to(
             hidden_states.dtype
         )
@@ -143,20 +151,24 @@ class Llama4Attention(nn.Module):
         self.hidden_size = hidden_size
         self.use_rope = int((layer_id + 1) % 4 != 0)
         self.use_qk_norm = config.use_qk_norm and self.use_rope
-        tp_size = get_tensor_model_parallel_world_size()
+
+        self.dp_size = get_attention_dp_size()
+        attn_tp_rank = get_attention_tp_rank()
+        attn_tp_size = get_attention_tp_size()
+
         self.total_num_heads = num_heads
-        assert self.total_num_heads % tp_size == 0
-        self.num_heads = self.total_num_heads // 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 >= 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 % 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 tp_size % self.total_num_kv_heads == 0
-        self.num_kv_heads = max(1, self.total_num_kv_heads // 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 = config.head_dim
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
@@ -183,6 +195,8 @@ class Llama4Attention(nn.Module):
             bias=bias,
             quant_config=quant_config,
             prefix=add_prefix("qkv_proj", prefix),
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
         )
 
         self.o_proj = RowParallelLinear(
@@ -191,6 +205,9 @@ class Llama4Attention(nn.Module):
             bias=bias_o_proj,
             quant_config=quant_config,
             prefix=add_prefix("o_proj", prefix),
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
+            reduce_results=False,
         )
         is_neox_style = True
         is_gguf = quant_config and quant_config.get_name() == "gguf"
@@ -223,9 +240,13 @@ class Llama4Attention(nn.Module):
     def _get_attn_scale(self, positions: torch.Tensor) -> torch.Tensor:
         floor = torch.floor((positions + 1.0) / self.floor_scale)
         attn_scale = torch.log(floor + 1.0) * self.attn_scale + 1.0
-
         return attn_scale.unsqueeze(-1)
 
+    @torch.compile(dynamic=True, backend=get_compiler_backend())
+    def _mul_attn_scale(self, positions, q):
+        attn_scale = self._get_attn_scale(positions)
+        return (q * attn_scale).to(q.dtype)
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -233,27 +254,29 @@ class Llama4Attention(nn.Module):
         forward_batch: ForwardBatch,
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
-        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+
+        qk, v = qkv.split([self.q_size + self.kv_size, self.kv_size], dim=-1)
 
         if self.rotary_emb is not None:
-            q, k = self.rotary_emb(positions, q, k)
+            q_view, k_view = qk.split([self.q_size, self.kv_size], dim=-1)
+            q_out_unused, k_out_unused = self.rotary_emb(positions, q_view, k_view)
+            assert (q_out_unused is q_view) and (k_out_unused is k_view)
+            del q_view, k_view, q_out_unused, k_out_unused
 
         if self.qk_norm is not None:
-            # TODO: support float
-            q = q.reshape(-1, self.head_dim).contiguous().bfloat16()
-            k = k.reshape(-1, self.head_dim).contiguous().bfloat16()
-            q = self.qk_norm(q).to(q.dtype)
-            k = self.qk_norm(k).to(k.dtype)
-            q = q.reshape(-1, self.q_size)
-            k = k.reshape(-1, self.kv_size)
+            # TODO there are still 2 redundant direct_copy_kernel_cuda for this `reshape` and (in attn backend) q.contiguous(), maybe we can fuse them later
+            qk = qk.reshape(-1, self.head_dim).contiguous().bfloat16()
+            qk = self.qk_norm(qk).to(torch.bfloat16)
+            qk = qk.reshape(-1, self.q_size + self.kv_size)
+
+        q, k = qk.split([self.q_size, self.kv_size], dim=-1)
 
         # We are applying temperature tuning (https://arxiv.org/abs/2501.19399) to NoPE layers, where
         # the inference-time temperature tuning function is customized to not affect short context
         # while working at very long context
         # https://arxiv.org/abs/2501.19399
         if self.attn_temperature_tuning and not self.use_rope:
-            attn_scale = self._get_attn_scale(positions)
-            q = (q * attn_scale).to(q.dtype)
+            q = self._mul_attn_scale(positions=positions, q=q)
 
         attn_output = self.attn(q, k, v, forward_batch)
         output, _ = self.o_proj(attn_output)
@@ -274,6 +297,9 @@ class Llama4DecoderLayer(nn.Module):
         rope_theta = config.rope_theta
         rope_scaling = config.rope_scaling
         max_position_embeddings = config.max_position_embeddings
+        self.dp_size = get_attention_dp_size()
+        self.attn_tp_size = get_attention_tp_size()
+        self.attn_tp_rank = get_attention_tp_rank()
 
         self.self_attn = Llama4Attention(
             config=config,
@@ -316,21 +342,58 @@ class Llama4DecoderLayer(nn.Module):
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        # Self Attention
-        if residual is None:
+        if hidden_states.shape[0] == 0:
             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,
-        )
+            # 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,
+            )
+
+        # Gather
+        if get_tensor_model_parallel_world_size() > 1:
+            # all gather and all reduce
+            if self.dp_size != 1:
+                if self.attn_tp_rank == 0:
+                    hidden_states += residual
+                hidden_states, local_hidden_states = (
+                    forward_batch.gathered_buffer,
+                    hidden_states,
+                )
+                dp_gather_partial(hidden_states, local_hidden_states, forward_batch)
+                dp_scatter(residual, hidden_states, forward_batch)
+                hidden_states = self.post_attention_layernorm(hidden_states)
+            else:
+                hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+                hidden_states, residual = self.post_attention_layernorm(
+                    hidden_states, residual
+                )
+        else:
+            hidden_states, residual = self.post_attention_layernorm(
+                hidden_states, residual
+            )
 
         # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
         hidden_states = self.feed_forward(hidden_states)
+
+        # TODO(ch-wan): ues reduce-scatter in MLP to avoid this scatter
+        # Scatter
+        if self.dp_size != 1:
+            # important: forward batch.gathered_buffer is used both after scatter and after gather.
+            # be careful about this!
+            hidden_states, global_hidden_states = (
+                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
+                hidden_states,
+            )
+            dp_scatter(hidden_states, global_hidden_states, forward_batch)
+
         return hidden_states, residual
 
 
@@ -350,13 +413,14 @@ class Llama4Model(nn.Module):
             config.hidden_size,
             quant_config=quant_config,
             prefix=add_prefix("embed_tokens", prefix),
+            enable_tp=not global_server_args_dict["enable_dp_attention"],
         )
         self.layers = make_layers(
             config.num_hidden_layers,
             lambda idx, prefix: Llama4DecoderLayer(
                 config=config, layer_id=idx, quant_config=quant_config, prefix=prefix
             ),
-            prefix="model.layers",
+            prefix=add_prefix("layers", prefix),
         )
 
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -385,7 +449,8 @@ class Llama4Model(nn.Module):
                 forward_batch,
                 residual,
             )
-        hidden_states, _ = self.norm(hidden_states, residual)
+        if not forward_batch.forward_mode.is_idle():
+            hidden_states, _ = self.norm(hidden_states, residual)
 
         if len(aux_hidden_states) == 0:
             return hidden_states
@@ -394,7 +459,6 @@ class Llama4Model(nn.Module):
 
 
 class Llama4ForCausalLM(LlamaForCausalLM):
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
@@ -408,6 +472,9 @@ class Llama4ForCausalLM(LlamaForCausalLM):
     ):
         super().__init__(config, quant_config, prefix)
 
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
     def _init_model(
         self,
         config: Llama4TextConfig,

sglang/srt/models/minicpm.py

@@ -146,6 +146,7 @@ class MiniCPMAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/minicpm3.py

@@ -192,6 +192,7 @@ class MiniCPM3Attention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_local_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 
@@ -343,6 +344,7 @@ class MiniCPM3AttentionMLA(nn.Module):
             num_kv_heads=1,
             layer_id=layer_id,
             v_head_dim=self.kv_lora_rank,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/mixtral.py

@@ -169,6 +169,7 @@ class MixtralAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/mixtral_quant.py

@@ -232,6 +232,7 @@ class MixtralAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/mllama.py

@@ -22,6 +22,7 @@ from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
     ColumnParallelLinear,
     QKVParallelLinear,
+    ReplicatedLinear,
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
@@ -184,6 +185,7 @@ class MllamaVisionEncoderLayer(nn.Module):
     def __init__(
         self,
         config: config_mllama.MllamaVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
         is_gated: bool = False,
         prefix: str = "",
     ):
@@ -199,14 +201,16 @@ class MllamaVisionEncoderLayer(nn.Module):
             self.num_attention_heads,
             self.hidden_size,
             use_qkv_parallel=True,
-            quant_config=None,
+            quant_config=quant_config,
             dropout=0.0,
             use_context_forward=False,
             softmax_in_single_precision=False,
             flatten_batch=False,
             prefix=add_prefix("self_attn", prefix),
         )
-        self.mlp = MllamaVisionMLP(config, prefix=add_prefix("mlp", prefix))
+        self.mlp = MllamaVisionMLP(
+            config, quant_config, prefix=add_prefix("mlp", prefix)
+        )
 
         self.input_layernorm = nn.LayerNorm(self.hidden_size, eps=config.norm_eps)
         self.post_attention_layernorm = nn.LayerNorm(
@@ -244,6 +248,7 @@ class MllamaVisionEncoder(nn.Module):
     def __init__(
         self,
         config: config_mllama.MllamaVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
         num_layers=32,
         is_gated=False,
         output_hidden_states=None,
@@ -254,7 +259,10 @@ class MllamaVisionEncoder(nn.Module):
         self.layers = nn.ModuleList(
             [
                 MllamaVisionEncoderLayer(
-                    config, is_gated, prefix=add_prefix(f"layers.{i}", prefix)
+                    config,
+                    quant_config,
+                    is_gated,
+                    prefix=add_prefix(f"layers.{i}", prefix),
                 )
                 for i in range(num_layers)
             ]
@@ -283,7 +291,12 @@ class MllamaVisionEncoder(nn.Module):
 
 
 class MllamaVisionModel(nn.Module):
-    def __init__(self, config: config_mllama.MllamaVisionConfig, prefix: str = ""):
+    def __init__(
+        self,
+        config: config_mllama.MllamaVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
         super().__init__()
         self.image_size = config.image_size
         self.patch_size = config.patch_size
@@ -320,6 +333,7 @@ class MllamaVisionModel(nn.Module):
         # encoders
         self.transformer = MllamaVisionEncoder(
             config,
+            quant_config,
             config.num_hidden_layers,
             is_gated=False,
             output_hidden_states=config.intermediate_layers_indices,
@@ -327,6 +341,7 @@ class MllamaVisionModel(nn.Module):
         )
         self.global_transformer = MllamaVisionEncoder(
             config,
+            quant_config,
             config.num_global_layers,
             is_gated=True,
             prefix=add_prefix("global_transformer", prefix),
@@ -535,6 +550,7 @@ class MllamaTextCrossAttention(nn.Module):
             self.num_local_key_value_heads,
             layer_id=layer_id,
             is_cross_attention=True,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 
@@ -764,6 +780,27 @@ class MllamaForCausalLM(nn.Module):
 
 
 class MllamaForConditionalGeneration(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    # in TP, these weights are partitioned along the column dimension (dim=-1)
+    column_parallel_weights_modules = [".down_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: config_mllama.MllamaConfig,
@@ -771,6 +808,7 @@ class MllamaForConditionalGeneration(nn.Module):
         prefix: str = "",
     ):
         super().__init__()
+        self.quant_config = quant_config
         self.vocab_size = config.text_config.vocab_size
         self.hidden_size = config.text_config.hidden_size
         self.max_num_tiles = config.vision_config.max_num_tiles
@@ -781,17 +819,21 @@ class MllamaForConditionalGeneration(nn.Module):
         self.image_size = config.vision_config.image_size
 
         self.vision_model = MllamaVisionModel(
-            config.vision_config, prefix=add_prefix("vision_model", prefix)
+            config.vision_config,
+            quant_config=quant_config,
+            prefix=add_prefix("vision_model", prefix),
         )
         self.language_model = MllamaForCausalLM(
             config.text_config,
             quant_config=quant_config,
             prefix=add_prefix("language_model", prefix),
         )
-        self.multi_modal_projector = nn.Linear(
+        self.multi_modal_projector = ReplicatedLinear(
             config.vision_config.vision_output_dim,
             config.text_config.hidden_size,
             bias=True,
+            quant_config=quant_config,
+            prefix="multi_modal_projector",
         )
         self.logits_processor = LogitsProcessor(config.text_config)
         self.capture_mode = False
@@ -958,7 +1000,9 @@ class MllamaForConditionalGeneration(nn.Module):
             cross_attention_states = self.vision_model(
                 batched_images, batched_ar_ids, batched_ar_mask
             )
-            cross_attention_states = self.multi_modal_projector(cross_attention_states)
+            cross_attention_states, _ = self.multi_modal_projector(
+                cross_attention_states
+            )
 
             bs, _, _, _, image_token_dim = cross_attention_states.shape
             cross_attention_states = cross_attention_states.view(
@@ -1012,7 +1056,6 @@ class MllamaForConditionalGeneration(nn.Module):
                 if "vision_model" in name:
                     # adapt to VisionAttention
                     name = name.replace("self_attn.o_proj", "self_attn.proj")
-
                 param = params_dict.pop(name)
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)

sglang/srt/models/mllama4.py

@@ -1,13 +1,19 @@
-# TODO: add Aapted from vllm/mllama4.py
 from collections.abc import Iterable
-from typing import Optional, Set, Tuple
+from typing import List, Optional, Set, Tuple
 
 import torch
 from torch import nn
-from transformers import Llama4Config
+from transformers import Llama4Config, Llama4VisionModel
+from transformers.models.llama4.modeling_llama4 import Llama4MultiModalProjector
 
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization import QuantizationConfig
+from sglang.srt.managers.mm_utils import (
+    MultiModalityDataPaddingPatternImageTokens,
+    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.utils import add_prefix
@@ -16,6 +22,7 @@ from sglang.srt.utils import add_prefix
 class Llama4ForConditionalGeneration(nn.Module):
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
     }
 
     def __init__(
@@ -28,6 +35,9 @@ class Llama4ForConditionalGeneration(nn.Module):
         self.config = config
         self.quant_config = quant_config
 
+        self.vision_model = Llama4VisionModel(config.vision_config)
+        self.multi_modal_projector = Llama4MultiModalProjector(config)
+
         # Initialize the language model
         from sglang.srt.models.llama4 import Llama4ForCausalLM
 
@@ -39,6 +49,29 @@ class Llama4ForConditionalGeneration(nn.Module):
 
         self.logits_processor = LogitsProcessor(config.text_config)
 
+    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 = MultiModalityDataPaddingPatternImageTokens(torch.tensor(im_token_id))
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
+
+    def get_image_feature(
+        self,
+        items: List[MultimodalDataItem],
+    ) -> torch.Tensor:
+        pixel_values = (
+            torch.concat([item.pixel_values for item in items])
+            .to(next(self.vision_model.parameters()).device)
+            .type(next(self.vision_model.parameters()).dtype)
+        )
+
+        image_outputs = self.vision_model(pixel_values, output_hidden_states=False)
+        image_features = image_outputs.last_hidden_state
+        vision_flat = image_features.view(-1, image_features.size(-1))
+        projected_vision_flat = self.multi_modal_projector(vision_flat)
+        return projected_vision_flat
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -47,7 +80,15 @@ class Llama4ForConditionalGeneration(nn.Module):
         **kwargs: object,
     ) -> torch.Tensor:
 
-        return self.language_model(input_ids, positions, forward_batch)
+        hs = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            language_model=self.language_model,
+            image_data_embedding_func=self.get_image_feature,
+            positions=positions,
+        )
+
+        return hs
 
     def permute_qk_weight_for_rotary(
         self,
@@ -96,18 +137,27 @@ class Llama4ForConditionalGeneration(nn.Module):
 
         num_experts = self.config.text_config.num_local_experts
 
-        for name, loaded_weight in weights:
-
-            if name.startswith("vision_model") or name.startswith(
-                "multi_modal_projector"
-            ):
-                continue
+        # Params for weights, fp8 weight scales, fp8 activation scales
+        # (param_name, weight_name, expert_id, shard_id)
+        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=num_experts,
+        )
 
-            name, loaded_weight = self.permute_qk_weight_for_rotary(name, loaded_weight)
+        for name, loaded_weight in weights:
+            if not "vision" in name:
+                name, loaded_weight = self.permute_qk_weight_for_rotary(
+                    name, loaded_weight
+                )
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
                     continue
+
+                if "vision" in name:
+                    continue
                 name = name.replace(weight_name, param_name)
                 param = params_dict[name]
                 weight_loader = param.weight_loader
@@ -115,31 +165,54 @@ class Llama4ForConditionalGeneration(nn.Module):
                 break
             else:
                 if ".experts" in name:
-                    if ".gate_up_proj" in name:
-                        name_list = [
-                            name.replace(".experts.gate_up_proj", ".experts.w13_weight")
-                        ] * 2
-                        loaded_weight_list = loaded_weight.chunk(2, dim=-1)
-                        shard_id_list = ["w1", "w3"]
-                    else:
-                        name_list = [
-                            name.replace(".experts.down_proj", ".experts.w2_weight")
-                        ]
-                        shard_id_list = ["w2"]
-                        loaded_weight_list = [loaded_weight]
-                    for name, loaded_weight, shard_id in zip(
-                        name_list, loaded_weight_list, shard_id_list
+                    # NOTE: llama4 fp8 has different weight format for experts
+                    if (
+                        "experts.gate_up_proj" not in name
+                        and "experts.down_proj" not in name
                     ):
-                        param = params_dict[name]
-                        weight_loader = param.weight_loader
-                        for expert_id in range(num_experts):
+                        for mapping in expert_params_mapping:
+                            param_name, weight_name, expert_id, shard_id = mapping
+                            if weight_name not in name:
+                                continue
+                            name = name.replace(weight_name, param_name)
+                            param = params_dict[name]
+                            weight_loader = param.weight_loader
                             weight_loader(
                                 param,
-                                loaded_weight[expert_id].T,
+                                loaded_weight,
                                 name,
                                 shard_id=shard_id,
                                 expert_id=expert_id,
                             )
+                            break
+                    else:
+                        if ".gate_up_proj" in name:
+                            name_list = [
+                                name.replace(
+                                    ".experts.gate_up_proj", ".experts.w13_weight"
+                                )
+                            ] * 2
+                            loaded_weight_list = loaded_weight.chunk(2, dim=-1)
+                            shard_id_list = ["w1", "w3"]
+                        else:
+                            name_list = [
+                                name.replace(".experts.down_proj", ".experts.w2_weight")
+                            ]
+                            shard_id_list = ["w2"]
+                            loaded_weight_list = [loaded_weight]
+                        for name, loaded_weight, shard_id in zip(
+                            name_list, loaded_weight_list, shard_id_list
+                        ):
+                            param = params_dict[name]
+                            weight_loader = param.weight_loader
+                            for expert_id in range(num_experts):
+                                weight_loader(
+                                    param,
+                                    loaded_weight[expert_id].T,
+                                    name,
+                                    shard_id=shard_id,
+                                    expert_id=expert_id,
+                                )
                 else:
                     # Skip loading extra bias for GPTQ models.
                     if name.endswith(".bias") and name not in params_dict:

sglang/srt/models/olmo.py

@@ -93,6 +93,7 @@ class OlmoAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )