sglang 0.2.11__py3-none-any.whl → 0.2.12__py3-none-any.whl

sglang/srt/model_executor/forward_batch_info.py

@@ -16,13 +16,17 @@ limitations under the License.
 """ModelRunner runs the forward passes of the models."""
 from dataclasses import dataclass
 from enum import IntEnum, auto
-from typing import List
+from typing import TYPE_CHECKING, List
 
 import numpy as np
 import torch
 
+from sglang.srt.managers.schedule_batch import ScheduleBatch
 from sglang.srt.mem_cache.memory_pool import BaseTokenToKVPool, ReqToTokenPool
 
+if TYPE_CHECKING:
+    from sglang.srt.model_executor.model_runner import ModelRunner
+
 
 class ForwardMode(IntEnum):
     # Prefill a new sequence. This is deprecated now. "EXTEND" covers this case.
@@ -39,25 +43,33 @@ class InputMetadata:
 
     forward_mode: ForwardMode
     batch_size: int
-    total_num_tokens: int
     req_pool_indices: torch.Tensor
     seq_lens: torch.Tensor
-    positions: torch.Tensor
     req_to_token_pool: ReqToTokenPool
     token_to_kv_pool: BaseTokenToKVPool
 
-    # For extend
-    extend_seq_lens: torch.Tensor
-    extend_start_loc: torch.Tensor
-    extend_no_prefix: bool
-
     # Output location of the KV cache
-    out_cache_loc: torch.Tensor = None
+    out_cache_loc: torch.Tensor
+
+    total_num_tokens: int = None
+
+    # Position information
+    positions: torch.Tensor = None
+
+    # For extend
+    extend_seq_lens: torch.Tensor = None
+    extend_start_loc: torch.Tensor = None
+    extend_no_prefix: bool = None
 
     # Output options
     return_logprob: bool = False
     top_logprobs_nums: List[int] = None
 
+    # For multimodal
+    pixel_values: List[torch.Tensor] = None
+    image_sizes: List[List[int]] = None
+    image_offsets: List[int] = None
+
     # Trition attention backend
     triton_max_seq_len: int = 0
     triton_max_extend_len: int = 0
@@ -70,107 +82,171 @@ class InputMetadata:
     flashinfer_decode_wrapper: "BatchDecodeWithPagedKVCacheWrapper" = None
     flashinfer_use_ragged: bool = False
 
-    @classmethod
-    def create(
-        cls,
-        model_runner,
-        forward_mode,
-        req_pool_indices,
-        seq_lens,
-        prefix_lens,
-        position_ids_offsets,
-        out_cache_loc,
-        top_logprobs_nums=None,
-        return_logprob=False,
-        skip_flashinfer_init=False,
-    ):
-        flashinfer_use_ragged = False
-        if not skip_flashinfer_init and not model_runner.server_args.disable_flashinfer:
-            if forward_mode != ForwardMode.DECODE and int(torch.sum(seq_lens)) > 4096:
-                flashinfer_use_ragged = True
-            init_flashinfer_args(
-                forward_mode,
-                model_runner,
-                req_pool_indices,
-                seq_lens,
-                prefix_lens,
-                model_runner.flashinfer_decode_wrapper,
-                flashinfer_use_ragged,
+    def init_multimuldal_info(self, batch: ScheduleBatch):
+        reqs = batch.reqs
+        self.pixel_values = [r.pixel_values for r in reqs]
+        self.image_sizes = [r.image_size for r in reqs]
+        self.image_offsets = [
+            (
+                (r.image_offset - len(r.prefix_indices))
+                if r.image_offset is not None
+                else 0
             )
+            for r in reqs
+        ]
 
-        batch_size = len(req_pool_indices)
+    def compute_positions(self, batch: ScheduleBatch):
+        position_ids_offsets = batch.position_ids_offsets
 
-        if forward_mode == ForwardMode.DECODE:
-            positions = ((seq_lens - 1) + position_ids_offsets).to(torch.int64)
-            extend_seq_lens = extend_start_loc = extend_no_prefix = None
-            if not model_runner.server_args.disable_flashinfer:
-                # This variable is not needed in this case,
-                # we do not compute it to make it compatbile with cuda graph.
-                total_num_tokens = None
+        if self.forward_mode == ForwardMode.DECODE:
+            if True:
+                self.positions = self.seq_lens - 1
             else:
-                total_num_tokens = int(torch.sum(seq_lens))
+                # Deprecated
+                self.positions = (self.seq_lens - 1) + position_ids_offsets
         else:
-            seq_lens_cpu = seq_lens.cpu().numpy()
-            prefix_lens_cpu = prefix_lens.cpu().numpy()
-            position_ids_offsets_cpu = position_ids_offsets.cpu().numpy()
-            positions = torch.tensor(
-                np.concatenate(
-                    [
-                        np.arange(
-                            prefix_lens_cpu[i] + position_ids_offsets_cpu[i],
-                            seq_lens_cpu[i] + position_ids_offsets_cpu[i],
-                        )
-                        for i in range(batch_size)
-                    ],
-                    axis=0,
-                ),
-                device="cuda",
-            )
-            extend_seq_lens = seq_lens - prefix_lens
-            extend_start_loc = torch.zeros_like(seq_lens)
-            extend_start_loc[1:] = torch.cumsum(extend_seq_lens[:-1], dim=0)
-            extend_no_prefix = torch.all(prefix_lens == 0)
-            total_num_tokens = int(torch.sum(seq_lens))
+            if True:
+                self.positions = torch.tensor(
+                    np.concatenate(
+                        [
+                            np.arange(len(req.prefix_indices), len(req.fill_ids))
+                            for req in batch.reqs
+                        ],
+                        axis=0,
+                    ),
+                    device="cuda",
+                )
+            else:
+                # Deprecated
+                position_ids_offsets_cpu = position_ids_offsets.cpu().numpy()
+                self.positions = torch.tensor(
+                    np.concatenate(
+                        [
+                            np.arange(
+                                len(req.prefix_indices) + position_ids_offsets_cpu[i],
+                                len(req.fill_ids) + position_ids_offsets_cpu[i],
+                            )
+                            for i, req in enumerate(batch.reqs)
+                        ],
+                        axis=0,
+                    ),
+                    device="cuda",
+                )
+
+        # Positions should be in long type
+        self.positions = self.positions.to(torch.int64)
+
+    def compute_extend_infos(self, batch: ScheduleBatch):
+        if self.forward_mode == ForwardMode.DECODE:
+            self.extend_seq_lens = self.extend_start_loc = self.extend_no_prefix = None
+        else:
+            extend_lens_cpu = [
+                len(r.fill_ids) - len(r.prefix_indices) for r in batch.reqs
+            ]
+            self.extend_seq_lens = torch.tensor(extend_lens_cpu, device="cuda")
+            self.extend_start_loc = torch.zeros_like(self.seq_lens)
+            self.extend_start_loc[1:] = torch.cumsum(self.extend_seq_lens[:-1], dim=0)
+            self.extend_no_prefix = all(len(r.prefix_indices) == 0 for r in batch.reqs)
 
+    @classmethod
+    def from_schedule_batch(
+        cls,
+        model_runner: "ModelRunner",
+        batch: ScheduleBatch,
+        forward_mode: ForwardMode,
+    ):
         ret = cls(
             forward_mode=forward_mode,
-            batch_size=batch_size,
-            total_num_tokens=total_num_tokens,
-            req_pool_indices=req_pool_indices,
-            seq_lens=seq_lens,
-            positions=positions,
+            batch_size=batch.batch_size(),
+            req_pool_indices=batch.req_pool_indices,
+            seq_lens=batch.seq_lens,
             req_to_token_pool=model_runner.req_to_token_pool,
             token_to_kv_pool=model_runner.token_to_kv_pool,
-            out_cache_loc=out_cache_loc,
-            extend_seq_lens=extend_seq_lens,
-            extend_start_loc=extend_start_loc,
-            extend_no_prefix=extend_no_prefix,
-            return_logprob=return_logprob,
-            top_logprobs_nums=top_logprobs_nums,
-            flashinfer_prefill_wrapper_ragged=model_runner.flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=model_runner.flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=model_runner.flashinfer_decode_wrapper,
-            flashinfer_use_ragged=flashinfer_use_ragged,
+            out_cache_loc=batch.out_cache_loc,
+            return_logprob=batch.return_logprob,
+            top_logprobs_nums=batch.top_logprobs_nums,
         )
 
+        ret.compute_positions(batch)
+
+        ret.compute_extend_infos(batch)
+
+        if (
+            forward_mode != ForwardMode.DECODE
+            or model_runner.server_args.disable_flashinfer
+        ):
+            ret.total_num_tokens = int(torch.sum(ret.seq_lens))
+
+        if forward_mode != ForwardMode.DECODE:
+            ret.init_multimuldal_info(batch)
+
+        prefix_lens = None
+        if forward_mode != ForwardMode.DECODE:
+            prefix_lens = torch.tensor(
+                [len(r.prefix_indices) for r in batch.reqs], device="cuda"
+            )
+
         if model_runner.server_args.disable_flashinfer:
-            (
-                ret.triton_max_seq_len,
-                ret.triton_max_extend_len,
-                ret.triton_start_loc,
-                ret.triton_prefix_lens,
-            ) = init_triton_args(forward_mode, seq_lens, prefix_lens)
+            ret.init_triton_args(batch, prefix_lens)
+
+        flashinfer_use_ragged = False
+        if not model_runner.server_args.disable_flashinfer:
+            if (
+                forward_mode != ForwardMode.DECODE
+                and int(torch.sum(ret.seq_lens)) > 4096
+            ):
+                flashinfer_use_ragged = True
+            ret.init_flashinfer_handlers(
+                model_runner, prefix_lens, flashinfer_use_ragged
+            )
 
         return ret
 
+    def init_triton_args(self, batch: ScheduleBatch, prefix_lens):
+        """Init auxiliary variables for triton attention backend."""
+        self.triton_max_seq_len = int(torch.max(self.seq_lens))
+        self.triton_prefix_lens = prefix_lens
+        self.triton_start_loc = torch.zeros_like(self.seq_lens, dtype=torch.int32)
+        self.triton_start_loc[1:] = torch.cumsum(self.seq_lens[:-1], dim=0)
+
+        if self.forward_mode == ForwardMode.DECODE:
+            self.triton_max_extend_len = None
+        else:
+            extend_seq_lens = self.seq_lens - prefix_lens
+            self.triton_max_extend_len = int(torch.max(extend_seq_lens))
+
+    def init_flashinfer_handlers(
+        self, model_runner, prefix_lens, flashinfer_use_ragged
+    ):
+        update_flashinfer_indices(
+            self.forward_mode,
+            model_runner,
+            self.req_pool_indices,
+            self.seq_lens,
+            prefix_lens,
+            flashinfer_use_ragged=flashinfer_use_ragged,
+        )
+
+        (
+            self.flashinfer_prefill_wrapper_ragged,
+            self.flashinfer_prefill_wrapper_paged,
+            self.flashinfer_decode_wrapper,
+            self.flashinfer_use_ragged,
+        ) = (
+            model_runner.flashinfer_prefill_wrapper_ragged,
+            model_runner.flashinfer_prefill_wrapper_paged,
+            model_runner.flashinfer_decode_wrapper,
+            flashinfer_use_ragged,
+        )
+
 
-def init_flashinfer_args(
+def update_flashinfer_indices(
     forward_mode,
     model_runner,
     req_pool_indices,
     seq_lens,
     prefix_lens,
-    flashinfer_decode_wrapper,
+    flashinfer_decode_wrapper=None,
     flashinfer_use_ragged=False,
 ):
     """Init auxiliary variables for FlashInfer attention backend."""
@@ -178,7 +254,6 @@ def init_flashinfer_args(
     num_kv_heads = model_runner.model_config.get_num_kv_heads(model_runner.tp_size)
     head_dim = model_runner.model_config.head_dim
     batch_size = len(req_pool_indices)
-    total_num_tokens = int(torch.sum(seq_lens))
 
     if flashinfer_use_ragged:
         paged_kernel_lens = prefix_lens
@@ -201,6 +276,10 @@ def init_flashinfer_args(
     kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda")
 
     if forward_mode == ForwardMode.DECODE:
+        # CUDA graph uses different flashinfer_decode_wrapper
+        if flashinfer_decode_wrapper is None:
+            flashinfer_decode_wrapper = model_runner.flashinfer_decode_wrapper
+
         flashinfer_decode_wrapper.end_forward()
         flashinfer_decode_wrapper.begin_forward(
             kv_indptr,
@@ -238,19 +317,3 @@ def init_flashinfer_args(
             head_dim,
             1,
         )
-
-
-def init_triton_args(forward_mode, seq_lens, prefix_lens):
-    """Init auxiliary variables for triton attention backend."""
-    batch_size = len(seq_lens)
-    max_seq_len = int(torch.max(seq_lens))
-    start_loc = torch.zeros((batch_size,), dtype=torch.int32, device="cuda")
-    start_loc[1:] = torch.cumsum(seq_lens[:-1], dim=0)
-
-    if forward_mode == ForwardMode.DECODE:
-        max_extend_len = None
-    else:
-        extend_seq_lens = seq_lens - prefix_lens
-        max_extend_len = int(torch.max(extend_seq_lens))
-
-    return max_seq_len, max_extend_len, start_loc, prefix_lens

sglang/srt/model_executor/model_runner.py

@@ -52,6 +52,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardMode, InputMetad
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import (
     get_available_gpu_memory,
+    is_generation_model,
     is_llama3_405b_fp8,
     is_multimodal_model,
     monkey_patch_vllm_dummy_weight_loader,
@@ -130,10 +131,12 @@ class ModelRunner:
             server_args.max_total_tokens,
         )
         self.init_cublas()
-        self.init_flash_infer()
+        self.init_flashinfer()
 
-        # Capture cuda graphs
-        self.init_cuda_graphs()
+        if self.is_generation:
+            # FIXME Currently, cuda graph only capture decode steps, which only exists in causal models
+            # Capture cuda graphs
+            self.init_cuda_graphs()
 
     def load_model(self):
         logger.info(
@@ -184,6 +187,10 @@ class ModelRunner:
             scheduler_config=None,
             cache_config=None,
         )
+        self.is_generation = is_generation_model(
+            self.model_config.hf_config.architectures
+        )
+
         logger.info(
             f"[gpu={self.gpu_id}] Load weight end. "
             f"type={type(self.model).__name__}, "
@@ -287,7 +294,7 @@ class ModelRunner:
         c = a @ b
         return c
 
-    def init_flash_infer(self):
+    def init_flashinfer(self):
         if self.server_args.disable_flashinfer:
             self.flashinfer_prefill_wrapper_ragged = None
             self.flashinfer_prefill_wrapper_paged = None
@@ -350,33 +357,18 @@ class ModelRunner:
         if self.cuda_graph_runner and self.cuda_graph_runner.can_run(len(batch.reqs)):
             return self.cuda_graph_runner.replay(batch)
 
-        input_metadata = InputMetadata.create(
-            self,
-            forward_mode=ForwardMode.DECODE,
-            req_pool_indices=batch.req_pool_indices,
-            seq_lens=batch.seq_lens,
-            prefix_lens=batch.prefix_lens,
-            position_ids_offsets=batch.position_ids_offsets,
-            out_cache_loc=batch.out_cache_loc,
-            top_logprobs_nums=batch.top_logprobs_nums,
-            return_logprob=batch.return_logprob,
+        input_metadata = InputMetadata.from_schedule_batch(
+            self, batch, ForwardMode.DECODE
         )
+
         return self.model.forward(
             batch.input_ids, input_metadata.positions, input_metadata
         )
 
     @torch.inference_mode()
     def forward_extend(self, batch: ScheduleBatch):
-        input_metadata = InputMetadata.create(
-            self,
-            forward_mode=ForwardMode.EXTEND,
-            req_pool_indices=batch.req_pool_indices,
-            seq_lens=batch.seq_lens,
-            prefix_lens=batch.prefix_lens,
-            position_ids_offsets=batch.position_ids_offsets,
-            out_cache_loc=batch.out_cache_loc,
-            top_logprobs_nums=batch.top_logprobs_nums,
-            return_logprob=batch.return_logprob,
+        input_metadata = InputMetadata.from_schedule_batch(
+            self, batch, forward_mode=ForwardMode.EXTEND
         )
         return self.model.forward(
             batch.input_ids, input_metadata.positions, input_metadata
@@ -384,24 +376,16 @@ class ModelRunner:
 
     @torch.inference_mode()
     def forward_extend_multi_modal(self, batch: ScheduleBatch):
-        input_metadata = InputMetadata.create(
-            self,
-            forward_mode=ForwardMode.EXTEND,
-            req_pool_indices=batch.req_pool_indices,
-            seq_lens=batch.seq_lens,
-            prefix_lens=batch.prefix_lens,
-            position_ids_offsets=batch.position_ids_offsets,
-            out_cache_loc=batch.out_cache_loc,
-            return_logprob=batch.return_logprob,
-            top_logprobs_nums=batch.top_logprobs_nums,
+        input_metadata = InputMetadata.from_schedule_batch(
+            self, batch, forward_mode=ForwardMode.EXTEND
         )
         return self.model.forward(
             batch.input_ids,
             input_metadata.positions,
             input_metadata,
-            batch.pixel_values,
-            batch.image_sizes,
-            batch.image_offsets,
+            input_metadata.pixel_values,
+            input_metadata.image_sizes,
+            input_metadata.image_offsets,
         )
 
     def forward(self, batch: ScheduleBatch, forward_mode: ForwardMode):
@@ -429,8 +413,10 @@ def import_model_classes():
                     entry, list
                 ):  # To support multiple model classes in one module
                     for tmp in entry:
+                        assert tmp.__name__ not in model_arch_name_to_cls
                         model_arch_name_to_cls[tmp.__name__] = tmp
                 else:
+                    assert entry.__name__ not in model_arch_name_to_cls
                     model_arch_name_to_cls[entry.__name__] = entry
 
             # compat: some models such as chatglm has incorrect class set in config.json
@@ -440,6 +426,7 @@ def import_model_classes():
             ):
                 for remap in module.EntryClassRemapping:
                     if isinstance(remap, tuple) and len(remap) == 2:
+                        assert remap[0] not in model_arch_name_to_cls
                         model_arch_name_to_cls[remap[0]] = remap[1]
 
     return model_arch_name_to_cls

sglang/srt/models/gemma2.py

@@ -38,7 +38,6 @@ from vllm.model_executor.layers.quantization.base_config import QuantizationConf
 # from vllm.model_executor.layers.rotary_embedding import GemmaRotaryEmbedding
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.model_executor.sampling_metadata import SamplingMetadata
 
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention

sglang/srt/models/internlm2.py

@@ -23,8 +23,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.activation import SiluAndMul
-from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
     QKVParallelLinear,
@@ -38,13 +36,14 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 )
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
 class InternLM2MLP(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
@@ -74,7 +73,6 @@ class InternLM2MLP(nn.Module):
 
 
 class InternLM2Attention(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
@@ -150,7 +148,6 @@ class InternLM2Attention(nn.Module):
 
 
 class InternLMDecoderLayer(nn.Module):
-
     def __init__(
         self,
         config: PretrainedConfig,
@@ -207,7 +204,6 @@ class InternLMDecoderLayer(nn.Module):
 
 
 class InternLM2Model(nn.Module):
-
     def __init__(
         self,
         config: PretrainedConfig,
@@ -254,7 +250,6 @@ class InternLM2Model(nn.Module):
 
 
 class InternLM2ForCausalLM(nn.Module):
-
     def __init__(
         self,
         config: PretrainedConfig,

sglang/srt/models/llama2.py

@@ -24,8 +24,6 @@ from torch import nn
 from transformers import LlamaConfig
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.activation import SiluAndMul
-from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
     QKVParallelLinear,
@@ -39,7 +37,9 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 )
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
-from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.logits_processor import LogitProcessorOutput, LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
@@ -310,7 +310,7 @@ class LlamaForCausalLM(nn.Module):
         positions: torch.Tensor,
         input_metadata: InputMetadata,
         input_embeds: torch.Tensor = None,
-    ) -> torch.Tensor:
+    ) -> LogitProcessorOutput:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
         return self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata

sglang/srt/models/llama_embedding.py

@@ -0,0 +1,88 @@
+from typing import Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.pooler import EmbeddingPoolerOutput, Pooler, PoolingType
+from sglang.srt.model_executor.model_runner import InputMetadata
+from sglang.srt.models.llama2 import LlamaForCausalLM, LlamaModel
+
+
+class LlamaEmbeddingModel(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config=None,
+        cache_config=None,
+        efficient_weight_load=False,
+    ) -> None:
+        super().__init__()
+        self.model = LlamaModel(config, quant_config=quant_config)
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> EmbeddingPoolerOutput:
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
+        return self.pooler(hidden_states, input_metadata)
+
+    def load_weights(
+        self, weights: Iterable[Tuple[str, torch.Tensor]], name=None, loaded_weight=None
+    ):
+        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.model.named_parameters())
+
+        def load_weights_per_param(name, loaded_weight):
+            if "rotary_emb.inv_freq" in name or "projector" in name:
+                return
+            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.
+                return
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if name.startswith("model.vision_tower") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    return
+                if name.startswith("model.vision_tower") and name not in params_dict:
+                    return
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+        if name is None or loaded_weight is None:
+            for name, loaded_weight in weights:
+                load_weights_per_param(name, loaded_weight)
+        else:
+            load_weights_per_param(name, loaded_weight)
+
+
+EntryClass = LlamaEmbeddingModel
+# compat: e5-mistral model.config class == MistralModel
+EntryClassRemapping = [("MistralModel", LlamaEmbeddingModel)]