release initial code

Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
Co-authored-by: parasol-aser <3848358+parasol-aser@users.noreply.github.com>
Co-authored-by: LiviaSun <33578456+ChuyueSun@users.noreply.github.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>

python/sglang/srt/models/llama2.py

@@ -0,0 +1,316 @@
+# Adapted from
+# https://github.com/vllm-project/vllm/blob/671af2b1c0b3ed6d856d37c21a561cc429a10701/vllm/model_executor/models/llama.py#L1
+"""Inference-only LLaMA model compatible with HuggingFace weights."""
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+from torch import nn
+from transformers import LlamaConfig
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    LinearMethodBase,
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size,
+)
+from vllm.model_executor.weight_utils import (
+    default_weight_loader,
+    hf_model_weights_iterator,
+)
+
+
+class LlamaMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size, hidden_size, bias=False, linear_method=linear_method
+        )
+        if hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Only silu is supported for now."
+            )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class LlamaAttention(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= 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
+        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)
+        self.head_dim = 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
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            linear_method=linear_method,
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class LlamaDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        layer_id: int = 0,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        self.self_attn = LlamaAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            linear_method=linear_method,
+        )
+        self.mlp = LlamaMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            linear_method=linear_method,
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        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,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class LlamaModel(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        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,
+        )
+        self.layers = nn.ModuleList(
+            [
+                LlamaDecoderLayer(config, i, linear_method)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        skip_embed: bool = False,
+    ) -> torch.Tensor:
+        if not skip_embed:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_ids
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class LlamaForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.model = LlamaModel(config, linear_method)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.logits_processor = LogitsProcessor(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        skip_embed: bool = False,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata, skip_embed)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    def load_weights(
+        self,
+        model_name_or_path: str,
+        cache_dir: Optional[str] = None,
+        load_format: str = "auto",
+        revision: Optional[str] = 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.named_parameters())
+        for name, loaded_weight in hf_model_weights_iterator(
+            model_name_or_path, cache_dir, load_format, revision
+        ):
+            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:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+            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
+                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:
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)

python/sglang/srt/models/llava.py

@@ -0,0 +1,213 @@
+"""Inference-only LLaVa model compatible with HuggingFace weights."""
+import json
+import os
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+from sglang.srt.managers.router.infer_batch import ForwardMode
+from sglang.srt.managers.router.model_runner import InputMetadata
+from sglang.srt.models.llama2 import LlamaForCausalLM
+from torch import nn
+from transformers import CLIPImageProcessor, CLIPVisionModel, LlavaConfig
+from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
+from vllm.model_executor.layers.linear import LinearMethodBase
+from vllm.model_executor.weight_utils import (
+    default_weight_loader,
+    hf_model_weights_iterator,
+)
+
+
+class LlavaLlamaForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: LlavaConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.vision_tower = None
+        self.config.vision_config.hidden_size = config.mm_hidden_size
+        self.config.text_config.hidden_size = config.hidden_size
+        self.multi_modal_projector = LlavaMultiModalProjector(config)
+        self.language_model = LlamaForCausalLM(config, linear_method)
+
+    def pad_input_ids(self, input_ids, pad_value):
+        pad_ids = pad_value * (
+            (self.image_feature_len + len(pad_value)) // len(pad_value)
+        )
+        offset = input_ids.index(self.config.image_token_index)
+        # old_len + pad_len - 1, because we need to remove image_token_id
+        new_input_ids = (
+            input_ids[:offset]
+            + pad_ids[: self.image_feature_len]
+            + input_ids[offset + 1 :]
+        )
+        return new_input_ids, offset
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        pixel_values: Optional[List[Optional[np.array]]] = None,
+        image_offsets: Optional[List[int]] = None,
+    ) -> torch.Tensor:
+        if input_metadata.forward_mode == ForwardMode.EXTEND:
+            bs = input_metadata.batch_size
+
+            # Embed text input
+            input_embeds = self.language_model.model.embed_tokens(input_ids)
+
+            # Embed vision input
+            need_vision = (
+                (positions[input_metadata.extend_start_loc] < self.image_feature_len)
+                .cpu()
+                .numpy()
+            )
+            # FIXME: We need to substract the length of the system prompt
+            has_pixel = np.array([pixel_values[i] is not None for i in range(bs)])
+            need_vision = need_vision & has_pixel
+
+            if need_vision.any():
+                pixel_values = torch.tensor(
+                    np.array([pixel_values[i] for i in range(bs) if need_vision[i]]),
+                    device=self.vision_tower.device,
+                )
+
+                image_outputs = self.vision_tower(
+                    pixel_values, output_hidden_states=True
+                )
+                # NOTE: This is not memory efficient. (output_hidden_states=True) will save all the hidden stated.
+
+                selected_image_feature = image_outputs.hidden_states[
+                    self.vision_feature_layer
+                ]
+                if self.vision_feature_select_strategy in ["default", "patch"]:
+                    selected_image_feature = selected_image_feature[:, 1:]
+                elif self.vision_feature_select_strategy == "full":
+                    selected_image_feature = selected_image_feature
+                else:
+                    raise ValueError(
+                        f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}"
+                    )
+                image_features = self.multi_modal_projector(selected_image_feature)
+
+                extend_start_loc_cpu = input_metadata.extend_start_loc.cpu().numpy()
+                pt = 0
+                for i in range(bs):
+                    if not need_vision[i]:
+                        continue
+
+                    start_idx = extend_start_loc_cpu[i]
+                    pad_len, pad_dim = image_features[pt].shape
+                    dim = input_embeds.shape[1]
+                    assert (
+                        pad_dim == dim
+                    ), "invalid pad_dim={}, input_embed_dim={}!".format(pad_dim, dim)
+                    # Fill in the placeholder for the image
+                    try:
+                        input_embeds[
+                            start_idx
+                            + image_offsets[i] : start_idx
+                            + image_offsets[i]
+                            + pad_len
+                        ] = image_features[pt]
+                    except RuntimeError as e:
+                        print(f"RuntimeError in llava image encoding: {e}")
+                        print(input_embeds.shape)
+                        print(start_idx, image_offsets[i])
+                    pt += 1
+
+            return self.language_model(
+                input_embeds, positions, input_metadata, skip_embed=True
+            )
+        elif input_metadata.forward_mode == ForwardMode.DECODE:
+            return self.language_model(
+                input_ids, positions, input_metadata, skip_embed=False
+            )
+
+    def load_weights(
+        self,
+        model_name_or_path: str,
+        cache_dir: Optional[str] = None,
+        load_format: str = "auto",
+        revision: Optional[str] = None,
+    ):
+        # load clip vision model by cfg['mm_vision_tower']:
+        #   huggingface_name or path_of_clip_relative_to_llava_model_dir
+        vision_path = self.config.mm_vision_tower
+        self.vision_tower = CLIPVisionModel.from_pretrained(
+            vision_path, torch_dtype=torch.float16
+        ).cuda()
+        self.vision_tower.eval()
+
+        self.vision_feature_layer = self.config.mm_vision_select_layer
+        self.vision_feature_select_strategy = self.config.mm_vision_select_feature
+        self.image_size = self.vision_tower.config.image_size
+        self.patch_size = self.vision_tower.config.patch_size
+        self.image_feature_len = int((self.image_size / self.patch_size) ** 2)
+        if self.vision_feature_select_strategy == "patch":
+            pass
+        elif self.vision_feature_select_strategy == "cls_patch":
+            self.image_feature_len += 1
+        else:
+            raise ValueError(f"Unexpected select feature: {self.select_feature}")
+
+        # load mm_projector
+        # TODO: support TP?
+        projector_weights = {
+            "model.mm_projector.0": "multi_modal_projector.linear_1",
+            "model.mm_projector.2": "multi_modal_projector.linear_2",
+        }
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in hf_model_weights_iterator(
+            model_name_or_path, cache_dir, load_format, revision
+        ):
+            # FIXME: why projector weights read two times?
+            if "projector" in name:
+                for weight_name, param_name in projector_weights.items():
+                    if weight_name in name:
+                        name = name.replace(weight_name, param_name)
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+        # load language model
+        self.language_model.load_weights(
+            model_name_or_path, cache_dir, load_format, revision
+        )
+
+        monkey_path_clip_vision_embed_forward()
+
+
+first_call = True
+
+
+def clip_vision_embed_forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+    batch_size = pixel_values.shape[0]
+
+    # Move this conv layer to CPU to avoid a bug in torch >= 2.1 on A10G.
+    global first_call
+    if first_call:
+        self.patch_embedding.cpu().float()
+        first_call = False
+    pixel_values = pixel_values.to(dtype=torch.float32, device="cpu")
+    patch_embeds = self.patch_embedding(pixel_values).cuda().half()
+
+    patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+    class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+    embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+    embeddings = embeddings + self.position_embedding(self.position_ids)
+    return embeddings
+
+
+def monkey_path_clip_vision_embed_forward():
+    import transformers
+
+    setattr(
+        transformers.models.clip.modeling_clip.CLIPVisionEmbeddings,
+        "forward",
+        clip_vision_embed_forward,
+    )

python/sglang/srt/models/mixtral.py

@@ -0,0 +1,378 @@
+# Adapted from
+# https://github.com/vllm-project/vllm/blob/d0215a58e78572d91dadafe9d832a2db89b09a13/vllm/model_executor/models/mixtral.py#L1
+"""Inference-only Mixtral model."""
+from typing import List, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+from torch import nn
+from transformers import MixtralConfig
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    LinearMethodBase,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.parallel_utils.communication_op import (
+    tensor_model_parallel_all_reduce,
+)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from vllm.model_executor.weight_utils import (
+    default_weight_loader,
+    hf_model_weights_iterator,
+)
+
+
+class MixtralMLP(nn.Module):
+    def __init__(
+        self,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size: int,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.num_experts = num_experts
+        self.ffn_dim = intermediate_size
+        self.hidden_dim = hidden_size
+
+        self.w1 = ReplicatedLinear(
+            self.hidden_dim, self.ffn_dim, bias=False, linear_method=linear_method
+        )
+        self.w2 = ReplicatedLinear(
+            self.ffn_dim, self.hidden_dim, bias=False, linear_method=linear_method
+        )
+        self.w3 = ReplicatedLinear(
+            self.hidden_dim, self.ffn_dim, bias=False, linear_method=linear_method
+        )
+
+        # TODO: Use vllm's SiluAndMul
+        self.act_fn = nn.SiLU()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        w1_out, _ = self.w1(hidden_states)
+        w1_out = self.act_fn(w1_out)
+        w3_out, _ = self.w3(hidden_states)
+        current_hidden_states = w1_out * w3_out
+        current_hidden_states, _ = self.w2(current_hidden_states)
+        return current_hidden_states
+
+
+class MixtralMoE(nn.Module):
+    def __init__(
+        self,
+        config: MixtralConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.rank = get_tensor_model_parallel_rank()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_total_experts = config.num_local_experts
+        self.top_k = config.num_experts_per_tok
+        if self.tp_size > self.num_total_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {self.num_total_experts}."
+            )
+        # Split experts equally between ranks
+        self.expert_indicies = np.array_split(
+            range(self.num_total_experts), self.tp_size
+        )[self.rank].tolist()
+        if not self.expert_indicies:
+            raise ValueError(f"Rank {self.rank} has no experts assigned to it.")
+
+        self.experts = nn.ModuleList(
+            [
+                MixtralMLP(
+                    self.num_total_experts,
+                    config.hidden_size,
+                    config.intermediate_size,
+                    linear_method=linear_method,
+                )
+                if idx in self.expert_indicies
+                else None
+                for idx in range(self.num_total_experts)
+            ]
+        )
+        self.gate = ReplicatedLinear(
+            config.hidden_size, self.num_total_experts, bias=False, linear_method=None
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        router_logits, _ = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(
+            routing_weights, self.top_k, dim=-1
+        )
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+
+        final_hidden_states = None
+        for expert_idx in self.expert_indicies:
+            expert_layer = self.experts[expert_idx]
+            expert_mask = selected_experts == expert_idx
+            expert_weights = (routing_weights * expert_mask).sum(dim=-1, keepdim=True)
+
+            current_hidden_states = expert_layer(hidden_states).mul_(expert_weights)
+            if final_hidden_states is None:
+                final_hidden_states = current_hidden_states
+            else:
+                final_hidden_states.add_(current_hidden_states)
+
+        return tensor_model_parallel_all_reduce(final_hidden_states)
+
+
+class MixtralAttention(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        max_position: int = 4096 * 32,
+        rope_theta: float = 10000,
+        linear_method: Optional[LinearMethodBase] = None,
+        sliding_window: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= 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
+        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)
+        self.head_dim = 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
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.sliding_window = sliding_window
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position,
+            base=int(self.rope_theta),
+            is_neox_style=True,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class MixtralDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: MixtralConfig,
+        layer_id: int = 0,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        # Requires transformers > 4.32.0
+        rope_theta = getattr(config, "rope_theta", 10000)
+        self.self_attn = MixtralAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            max_position=config.max_position_embeddings,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            sliding_window=config.sliding_window,
+            linear_method=linear_method,
+        )
+        self.block_sparse_moe = MixtralMoE(config=config, linear_method=linear_method)
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[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,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.block_sparse_moe(hidden_states)
+        return hidden_states, residual
+
+
+class MixtralModel(nn.Module):
+    def __init__(
+        self,
+        config: MixtralConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        # config.num_hidden_layers=16
+        self.layers = nn.ModuleList(
+            [
+                MixtralDecoderLayer(config, i, linear_method=linear_method)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        skip_embed: bool = False,
+    ) -> torch.Tensor:
+        if not skip_embed:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_ids
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions, hidden_states, input_metadata, residual
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class MixtralForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: MixtralConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.model = MixtralModel(config, linear_method)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.logits_processor = LogitsProcessor(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        skip_embed: bool = False,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata, skip_embed)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    def load_weights(
+        self,
+        model_name_or_path: str,
+        cache_dir: Optional[str] = None,
+        load_format: str = "auto",
+        revision: Optional[str] = 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"),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in hf_model_weights_iterator(
+            model_name_or_path, cache_dir, load_format, revision, fall_back_to_pt=False
+        ):
+            if "rotary_emb.inv_freq" in name:
+                continue
+            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
+                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:
+                    continue
+                # Skip experts that are not assigned to this worker.
+                if "block_sparse_moe.experts." in name 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)