merge main

vllm/model_executor/models/gpt_j.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gptj/modeling_gptj.py
+# Copyright 2023 The vLLM team.
+# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only GPT-J model compatible with HuggingFace weights."""
+from typing import List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import GPTJConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               LinearMethodBase,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class GPTJAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTJConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.total_num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.total_num_heads
+
+        self.qkv_proj = QKVParallelLinear(
+            config.hidden_size,
+            self.head_size,
+            self.total_num_heads,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.out_proj = RowParallelLinear(
+            config.hidden_size,
+            config.hidden_size,
+            bias=False,
+            linear_method=linear_method,
+        )
+
+        tp_world_size = get_tensor_model_parallel_world_size()
+        assert self.total_num_heads % tp_world_size == 0
+        self.num_heads = self.total_num_heads // tp_world_size
+
+        scaling = self.head_size**-0.5
+        assert getattr(config, "rotary", True)
+        assert config.rotary_dim % 2 == 0
+        rope_theta = getattr(config, "rope_theta", 10000)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.rotary_emb = get_rope(
+            self.head_size,
+            rotary_dim=config.rotary_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            is_neox_style=False,
+        )
+        self.attn = PagedAttention(self.num_heads, self.head_size, scaling)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        q, k = self.rotary_emb(position_ids, q, k)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        attn_output, _ = self.out_proj(attn_output)
+        return attn_output
+
+
+class GPTJMLP(nn.Module):
+
+    def __init__(
+        self,
+        intermediate_size: int,
+        config: GPTJConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        hidden_size = config.n_embd
+        self.fc_in = ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            linear_method=linear_method,
+        )
+        self.fc_out = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            linear_method=linear_method,
+        )
+        quant_config = getattr(linear_method, "quant_config", None)
+        self.act = get_act_fn(config.activation_function, quant_config,
+                              intermediate_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.fc_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.fc_out(hidden_states)
+        return hidden_states
+
+
+class GPTJBlock(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTJConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        inner_dim = 4 * config.n_embd if config.n_inner is None else config.n_inner
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = GPTJAttention(config, linear_method)
+        self.mlp = GPTJMLP(inner_dim, config, linear_method)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_output = self.attn(
+            position_ids=position_ids,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            input_metadata=input_metadata,
+        )
+        mlp_output = self.mlp(hidden_states)
+        hidden_states = attn_output + mlp_output + residual
+        return hidden_states
+
+
+class GPTJModel(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTJConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.n_embd
+        self.wte = VocabParallelEmbedding(
+            config.vocab_size,
+            self.embed_dim,
+        )
+        self.h = nn.ModuleList(
+            [GPTJBlock(config, linear_method) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.wte(input_ids)
+        for i in range(len(self.h)):
+            layer = self.h[i]
+            hidden_states = layer(
+                position_ids,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+            )
+        hidden_states = self.ln_f(hidden_states)
+        return hidden_states
+
+
+class GPTJForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTJConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        assert not config.tie_word_embeddings
+        self.transformer = GPTJModel(config, linear_method)
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.n_embd,
+            bias=True,
+        )
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.transformer(input_ids, positions, kv_caches,
+                                         input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
+                                   sampling_metadata, self.lm_head.bias)
+        return next_tokens
+
+    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 "attn.bias" in name or "attn.masked_bias" 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
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)

vllm/model_executor/models/gpt_neox.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gpt_neox/modeling_gpt_neox.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only GPT-NeoX model compatible with HuggingFace weights."""
+from typing import List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import GPTNeoXConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               LinearMethodBase,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class GPTNeoXAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTNeoXConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.total_num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.total_num_heads
+
+        tensor_model_parallel_world_size = (
+            get_tensor_model_parallel_world_size())
+        assert self.total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = (self.total_num_heads //
+                          tensor_model_parallel_world_size)
+
+        self.query_key_value = QKVParallelLinear(
+            config.hidden_size,
+            self.head_size,
+            self.total_num_heads,
+            linear_method=linear_method,
+        )
+        self.dense = RowParallelLinear(
+            config.hidden_size,
+            config.hidden_size,
+            linear_method=linear_method,
+        )
+        scaling = self.head_size**-0.5
+        rotary_dim = int(self.head_size * config.rotary_pct)
+        assert rotary_dim % 2 == 0
+        rope_theta = getattr(config, "rope_theta", 10000)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.rotary_emb = get_rope(
+            self.head_size,
+            rotary_dim=rotary_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+        )
+        self.attn = PagedAttention(self.num_heads, self.head_size, scaling)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.query_key_value(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        q, k = self.rotary_emb(position_ids, q, k)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.dense(attn_output)
+        return output
+
+
+class GPTNeoXMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTNeoXConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.dense_h_to_4h = ColumnParallelLinear(
+            config.hidden_size,
+            config.intermediate_size,
+            linear_method=linear_method,
+        )
+        self.dense_4h_to_h = RowParallelLinear(
+            config.intermediate_size,
+            config.hidden_size,
+            linear_method=linear_method,
+        )
+        quant_config = getattr(linear_method, "quant_config", None)
+        self.act = get_act_fn(config.hidden_act, quant_config,
+                              config.intermediate_size)
+
+    def forward(self, hidden_states):
+        hidden_states, _ = self.dense_h_to_4h(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.dense_4h_to_h(hidden_states)
+        return hidden_states
+
+
+class GPTNeoXLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTNeoXConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.use_parallel_residual = config.use_parallel_residual
+        self.input_layernorm = nn.LayerNorm(config.hidden_size,
+                                            eps=config.layer_norm_eps)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
+                                                     eps=config.layer_norm_eps)
+        self.attention = GPTNeoXAttention(config, linear_method)
+        self.mlp = GPTNeoXMLP(config, linear_method)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        attn_input = self.input_layernorm(hidden_states)
+        attn_output = self.attention(
+            position_ids=position_ids,
+            hidden_states=attn_input,
+            kv_cache=kv_cache,
+            input_metadata=input_metadata,
+        )
+
+        if self.use_parallel_residual:
+            # pseudocode:
+            # x = x + attn(ln1(x)) + mlp(ln2(x))
+            mlp_input = self.post_attention_layernorm(hidden_states)
+            mlp_output = self.mlp(mlp_input)
+            hidden_states = mlp_output + attn_output + hidden_states
+        else:
+            # pseudocode:
+            # x = x + attn(ln1(x))
+            # x = x + mlp(ln2(x))
+            attn_output = attn_output + hidden_states
+            mlp_input = self.post_attention_layernorm(attn_output)
+            mlp_output = self.mlp(mlp_input)
+            hidden_states = mlp_output + attn_output
+        return hidden_states
+
+
+class GPTNeoXModel(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTNeoXConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+
+        self.embed_in = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList([
+            GPTNeoXLayer(config, linear_method)
+            for _ in range(config.num_hidden_layers)
+        ])
+        self.final_layer_norm = nn.LayerNorm(config.hidden_size,
+                                             eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_in(input_ids)
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states = layer(
+                position_ids,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+            )
+        hidden_states = self.final_layer_norm(hidden_states)
+        return hidden_states
+
+
+class GPTNeoXForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.gpt_neox = GPTNeoXModel(config, linear_method)
+        self.embed_out = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.gpt_neox(input_ids, positions, kv_caches,
+                                      input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.embed_out.weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    def load_weights(self,
+                     model_name_or_path: str,
+                     cache_dir: Optional[str] = None,
+                     load_format: str = "auto",
+                     revision: Optional[str] = None):
+        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 ("attention.bias" in name or "attention.masked_bias" in name
+                    or "rotary_emb.inv_freq" in name):
+                continue
+            param = params_dict[name]
+
+            if "query_key_value" in name:
+                # NOTE: GPT-NeoX's fused QKV's output_dim has the shape of
+                # (num_heads * 3 * head_size), while the
+                # required shape is (3 * num_heads * head_size).
+                # Thus, we need weight conversion.
+                output_dim = getattr(param, "output_dim", None)
+                num_heads = self.config.num_attention_heads
+                if output_dim is not None:
+                    loaded_weight_shape = loaded_weight.shape
+                    loaded_weight = loaded_weight.view(
+                        loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
+                        loaded_weight_shape[output_dim + 1:])
+                    loaded_weight = loaded_weight.transpose(
+                        output_dim, output_dim + 1)
+                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)
+
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)

vllm/model_executor/models/internlm.py

+# -*- coding: utf-8 -*-
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.attention import PagedAttention
+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.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class InternLMMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        linear_method: Optional[LinearMethodBase] = 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 InternLMAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        bias: bool,
+        rope_theta: float = 10000,
+        max_position_embeddings: int = 8192,
+        linear_method: Optional[LinearMethodBase] = None,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        tensor_model_parallel_world_size = (
+            get_tensor_model_parallel_world_size())
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = (self.total_num_heads //
+                          tensor_model_parallel_world_size)
+        self.head_dim = hidden_size // self.total_num_heads
+        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,
+            bias=bias,
+            linear_method=linear_method,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=bias,
+            linear_method=linear_method,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=self.max_position_embeddings,
+            base=self.rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = PagedAttention(self.num_heads, self.head_dim, self.scaling)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class InternLMDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.self_attn = InternLMAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            bias=config.bias,
+            rope_theta=rope_theta,
+            max_position_embeddings=max_position_embeddings,
+            linear_method=linear_method,
+            rope_scaling=getattr(config, "rope_scaling", None),
+        )
+        self.mlp = InternLMMLP(
+            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,
+        kv_cache: KVCache,
+        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,
+            kv_cache=kv_cache,
+            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 InternLMModel(nn.Module):
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        vocab_size = ((config.vocab_size + 63) // 64) * 64
+        self.embed_tokens = VocabParallelEmbedding(
+            vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList([
+            InternLMDecoderLayer(config, linear_method)
+            for _ 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,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class InternLMForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.model = InternLMModel(config, linear_method)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    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:
+                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)

vllm/model_executor/models/llama.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only LLaMA model compatible with HuggingFace weights."""
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.attention import PagedAttention
+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.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+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,
+        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 = PagedAttention(self.num_heads,
+                                   self.head_dim,
+                                   self.scaling,
+                                   num_kv_heads=self.num_kv_heads)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        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)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class LlamaDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+        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,
+            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,
+        kv_cache: KVCache,
+        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,
+            kv_cache=kv_cache,
+            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, linear_method)
+            for _ 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,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i],
+                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.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    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:
+                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)

vllm/model_executor/models/mistral.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only Mistral model compatible with HuggingFace weights."""
+from typing import List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import MistralConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.attention import PagedAttention
+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.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class MistralMLP(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 MistralAttention(nn.Module):
+
+    def __init__(self,
+                 hidden_size: int,
+                 num_heads: int,
+                 num_kv_heads: int,
+                 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=self.rope_theta,
+        )
+        self.attn = PagedAttention(self.num_heads,
+                                   self.head_dim,
+                                   self.scaling,
+                                   num_kv_heads=self.num_kv_heads,
+                                   sliding_window=self.sliding_window)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        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)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class MistralDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: MistralConfig,
+        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 = MistralAttention(
+            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,
+            rope_theta=rope_theta,
+            linear_method=linear_method,
+            sliding_window=config.sliding_window)
+        self.mlp = MistralMLP(
+            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,
+        kv_cache: KVCache,
+        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,
+            kv_cache=kv_cache,
+            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 MistralModel(nn.Module):
+
+    def __init__(
+        self,
+        config: MistralConfig,
+        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([
+            MistralDecoderLayer(config, linear_method)
+            for _ 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,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class MistralForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: MistralConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.model = MistralModel(config, linear_method)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    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:
+                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)

vllm/model_executor/models/mixtral.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only Mixtral model."""
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+import torch
+import torch.nn.functional as F
+
+from torch import nn
+from transformers import MixtralConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (LinearMethodBase,
+                                               ReplicatedLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+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.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+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:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        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).view(
+            batch_size, sequence_length, hidden_dim)
+
+
+class MixtralAttention(nn.Module):
+
+    def __init__(self,
+                 hidden_size: int,
+                 num_heads: int,
+                 num_kv_heads: int,
+                 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 = PagedAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            sliding_window=self.sliding_window,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        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)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class MixtralDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: MixtralConfig,
+        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,
+            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,
+        kv_cache: KVCache,
+        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,
+            kv_cache=kv_cache,
+            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,
+        )
+        self.layers = nn.ModuleList([
+            MixtralDecoderLayer(config, linear_method=linear_method)
+            for _ 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,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> SamplerOutput:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(positions, hidden_states,
+                                            kv_caches[i], 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.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: Optional[torch.Tensor],
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    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)

vllm/model_executor/models/mpt.py

+# coding=utf-8
+# Adapted from https://huggingface.co/mosaicml/mpt-7b/tree/main
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               LinearMethodBase,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+from vllm.transformers_utils.configs.mpt import MPTConfig
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+def _get_alibi_slopes(
+    total_num_heads: int,
+    alibi_bias_max: int,
+) -> torch.Tensor:
+    next_power_of_2 = 2**math.ceil(math.log2(total_num_heads))
+    m = torch.arange(1, next_power_of_2 + 1, dtype=torch.float32)
+    m = m.mul(alibi_bias_max / next_power_of_2)
+    slopes = 1.0 / torch.pow(2, m)
+    if next_power_of_2 != total_num_heads:
+        slopes = torch.concat([slopes[1::2], slopes[::2]])[:total_num_heads]
+    return slopes
+
+
+class MPTAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: MPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.d_model = config.d_model
+        self.total_num_heads = config.n_heads
+        self.head_dim = self.d_model // self.total_num_heads
+        self.clip_qkv = config.attn_config["clip_qkv"]
+        self.qk_ln = config.attn_config["qk_ln"]
+        self.alibi_bias_max = config.attn_config["alibi_bias_max"]
+        if "kv_n_heads" in config.attn_config:
+            self.total_num_kv_heads = config.attn_config['kv_n_heads']
+        else:
+            self.total_num_kv_heads = self.total_num_heads
+        assert not config.attn_config["prefix_lm"]
+        assert config.attn_config["alibi"]
+
+        # pylint: disable=invalid-name
+        self.Wqkv = QKVParallelLinear(
+            self.d_model,
+            self.d_model // self.total_num_heads,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=not config.no_bias,
+            linear_method=linear_method,
+        )
+        if self.qk_ln:
+            self.q_ln = nn.LayerNorm(self.d_model)
+            self.k_ln = nn.LayerNorm(self.d_model)
+        self.out_proj = RowParallelLinear(
+            self.d_model,
+            self.d_model,
+            bias=not config.no_bias,
+            linear_method=linear_method,
+        )
+
+        tp_world_size = get_tensor_model_parallel_world_size()
+        assert self.total_num_heads % tp_world_size == 0
+        self.num_heads = self.total_num_heads // tp_world_size
+
+        if self.total_num_kv_heads >= tp_world_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_world_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_world_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_world_size)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        # Create the alibi slopes and slice them.
+        tp_rank = get_tensor_model_parallel_rank()
+        head_start = tp_rank * self.num_heads
+        head_end = (tp_rank + 1) * self.num_heads
+        alibi_slopes = _get_alibi_slopes(self.total_num_heads,
+                                         self.alibi_bias_max)
+        alibi_slopes = alibi_slopes[head_start:head_end].tolist()
+
+        self.head_dim = self.d_model // self.total_num_heads
+        scaling = self.head_dim**-0.5
+        self.attn = PagedAttention(self.num_heads,
+                                   self.head_dim,
+                                   scaling,
+                                   alibi_slopes=alibi_slopes,
+                                   num_kv_heads=self.num_kv_heads)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        del position_ids  # unused.
+        qkv, _ = self.Wqkv(hidden_states)
+        if self.clip_qkv is not None:
+            qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        if self.qk_ln:
+            q = self.q_ln(q)
+            k = self.k_ln(k)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.out_proj(attn_output)
+        return output
+
+
+class MPTMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: MPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        hidden_size = config.d_model
+        expansion_ratio = config.expansion_ratio
+        intermediate_size = expansion_ratio * hidden_size
+        self.up_proj = ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            bias=not config.no_bias,
+            linear_method=linear_method,
+        )
+        quant_config = getattr(linear_method, "quant_config", None)
+        self.act = get_act_fn("gelu", quant_config, intermediate_size)
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=not config.no_bias,
+            linear_method=linear_method,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x, _ = self.up_proj(x)
+        x = self.act(x)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class MPTBlock(nn.Module):
+
+    def __init__(
+        self,
+        config: MPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        hidden_size = config.d_model
+        self.norm_1 = nn.LayerNorm(hidden_size)
+        self.attn = MPTAttention(config, linear_method)
+        self.norm_2 = nn.LayerNorm(hidden_size)
+        self.ffn = MPTMLP(config, linear_method)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        x = self.norm_1(hidden_states)
+        x = self.attn(
+            position_ids=position_ids,
+            hidden_states=x,
+            kv_cache=kv_cache,
+            input_metadata=input_metadata,
+        )
+        hidden_states = hidden_states + x
+        x = self.norm_2(hidden_states)
+        x = self.ffn(x)
+        hidden_states = hidden_states + x
+        return hidden_states
+
+
+class MPTModel(nn.Module):
+
+    def __init__(
+        self,
+        config: MPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        assert config.embedding_fraction == 1.0
+        assert config.norm_type == "low_precision_layernorm"
+
+        self.wte = VocabParallelEmbedding(
+            config.vocab_size,
+            config.d_model,
+        )
+        self.blocks = nn.ModuleList(
+            [MPTBlock(config, linear_method) for _ in range(config.n_layers)])
+        self.norm_f = nn.LayerNorm(config.d_model)
+        if config.no_bias:
+            for module in self.modules():
+                if hasattr(module, "bias") and isinstance(
+                        module.bias, nn.Parameter):
+                    # Remove the bias term in Linear and LayerNorm.
+                    module.register_parameter("bias", None)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.wte(input_ids)
+        for i in range(len(self.blocks)):
+            block = self.blocks[i]
+            hidden_states = block(
+                position_ids,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+            )
+        hidden_states = self.norm_f(hidden_states)
+        return hidden_states
+
+
+class MPTForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: MPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        assert config.tie_word_embeddings
+        self.linear_method = linear_method
+
+        self.transformer = MPTModel(config, linear_method)
+        self.lm_head_weight = self.transformer.wte.weight
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.transformer(input_ids, positions, kv_caches,
+                                         input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head_weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    def load_weights(self,
+                     model_name_or_path: str,
+                     cache_dir: Optional[str] = None,
+                     load_format: str = "auto",
+                     revision: Optional[str] = None):
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        for name, loaded_weight in hf_model_weights_iterator(
+                model_name_or_path, cache_dir, load_format, revision):
+            # 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)

vllm/model_executor/models/opt.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/opt/modeling_opt.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights
+# reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only OPT model compatible with HuggingFace weights."""
+from typing import List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import OPTConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               LinearMethodBase,
+                                               QKVParallelLinear,
+                                               ReplicatedLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class OPTLearnedPositionalEmbedding(nn.Embedding):
+
+    def __init__(self, num_embeddings: int, embedding_dim: int):
+        # OPT is set up so that if padding_idx is specified then offset the
+        # embedding ids by 2 and adjust num_embeddings appropriately. Other
+        # models don't have this hack
+        self.offset = 2
+        super().__init__(num_embeddings + self.offset, embedding_dim)
+
+    def forward(self, positions: torch.Tensor):
+        return super().forward(positions + self.offset)
+
+
+class OPTAttention(nn.Module):
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        bias: bool = True,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.embed_dim = embed_dim
+        tensor_model_parallel_world_size = (
+            get_tensor_model_parallel_world_size())
+        total_num_heads = num_heads
+        assert num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = total_num_heads // tensor_model_parallel_world_size
+        self.head_dim = embed_dim // total_num_heads
+        self.scaling = self.head_dim**-0.5
+
+        self.qkv_proj = QKVParallelLinear(
+            embed_dim,
+            self.head_dim,
+            total_num_heads,
+            bias=bias,
+            linear_method=linear_method,
+        )
+        self.out_proj = RowParallelLinear(
+            embed_dim,
+            embed_dim,
+            bias=bias,
+            linear_method=linear_method,
+        )
+        self.attn = PagedAttention(self.num_heads,
+                                   self.head_dim,
+                                   scale=self.scaling)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        key_cache, value_cache = kv_cache
+        attn_output = self.attn(q, k, v, key_cache, value_cache,
+                                input_metadata)
+        output, _ = self.out_proj(attn_output)
+        return output
+
+
+class OPTDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: OPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.self_attn = OPTAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.num_attention_heads,
+            bias=config.enable_bias,
+            linear_method=linear_method,
+        )
+        self.do_layer_norm_before = config.do_layer_norm_before
+
+        self.self_attn_layer_norm = nn.LayerNorm(
+            self.embed_dim,
+            elementwise_affine=config.layer_norm_elementwise_affine)
+        self.fc1 = ColumnParallelLinear(
+            self.embed_dim,
+            config.ffn_dim,
+            bias=config.enable_bias,
+            linear_method=linear_method,
+        )
+        quant_config = getattr(linear_method, "quant_config", None)
+        self.activation_fn = get_act_fn(config.activation_function,
+                                        quant_config, config.ffn_dim)
+        self.fc2 = RowParallelLinear(
+            config.ffn_dim,
+            self.embed_dim,
+            bias=config.enable_bias,
+            linear_method=linear_method,
+        )
+        self.final_layer_norm = nn.LayerNorm(
+            self.embed_dim,
+            elementwise_affine=config.layer_norm_elementwise_affine)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        # Self Attention
+        residual = hidden_states
+        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
+        if self.do_layer_norm_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states = self.self_attn(hidden_states=hidden_states,
+                                       kv_cache=kv_cache,
+                                       input_metadata=input_metadata)
+        hidden_states = residual + hidden_states
+        # 350m applies layer norm AFTER attention
+        if not self.do_layer_norm_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        residual = hidden_states
+        # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
+        if self.do_layer_norm_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states, _ = self.fc2(hidden_states)
+        hidden_states = residual + hidden_states
+        # 350m applies layer norm AFTER attention
+        if not self.do_layer_norm_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+        return hidden_states
+
+
+class OPTDecoder(nn.Module):
+
+    def __init__(
+        self,
+        config: OPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.word_embed_proj_dim,
+        )
+        # Positional embeddings are replicated (not sharded).
+        self.embed_positions = OPTLearnedPositionalEmbedding(
+            config.max_position_embeddings, config.hidden_size)
+
+        # Project out & in will be replicated if they exist.
+        if config.word_embed_proj_dim != config.hidden_size:
+            self.project_out = ReplicatedLinear(config.hidden_size,
+                                                config.word_embed_proj_dim,
+                                                bias=False,
+                                                linear_method=linear_method)
+        else:
+            self.project_out = None
+
+        if config.word_embed_proj_dim != config.hidden_size:
+            self.project_in = ReplicatedLinear(config.word_embed_proj_dim,
+                                               config.hidden_size,
+                                               bias=False,
+                                               linear_method=linear_method)
+        else:
+            self.project_in = None
+
+        # Note that the only purpose of `config._remove_final_layer_norm` is to
+        # keep backward compatibility with checkpoints that have been fine-tuned
+        # before transformers v4.20.1
+        # see https://github.com/facebookresearch/metaseq/pull/164
+        if config.do_layer_norm_before and not config._remove_final_layer_norm:
+            self.final_layer_norm = nn.LayerNorm(
+                config.hidden_size,
+                elementwise_affine=config.layer_norm_elementwise_affine)
+        else:
+            self.final_layer_norm = None
+
+        self.layers = nn.ModuleList([
+            OPTDecoderLayer(config, linear_method)
+            for _ in range(config.num_hidden_layers)
+        ])
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        inputs_embeds = self.embed_tokens(input_ids)
+        pos_embeds = self.embed_positions(positions)
+        if self.project_in is not None:
+            inputs_embeds, _ = self.project_in(inputs_embeds)
+        hidden_states = inputs_embeds + pos_embeds
+
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states = layer(hidden_states, kv_caches[i], input_metadata)
+
+        if self.final_layer_norm is not None:
+            hidden_states = self.final_layer_norm(hidden_states)
+        if self.project_out is not None:
+            hidden_states, _ = self.project_out(hidden_states)
+        return hidden_states
+
+
+class OPTModel(nn.Module):
+
+    def __init__(
+        self,
+        config: OPTConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.decoder = OPTDecoder(config, linear_method)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        return self.decoder(input_ids, positions, kv_caches, input_metadata)
+
+
+class OPTForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.model = OPTModel(config, linear_method)
+        self.lm_head_weight = self.model.decoder.embed_tokens.weight
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head_weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    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(remove_duplicate=False))
+        for name, loaded_weight in hf_model_weights_iterator(
+                model_name_or_path, cache_dir, load_format, revision):
+            if "lm_head.weight" in name:
+                continue
+            if name.startswith("decoder."):
+                name = "model." + name
+
+            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)

vllm/model_executor/models/phi_1_5.py

+# coding=utf-8
+# Adapted from
+# https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_phi.py
+# Copyright 2023 The vLLM team.
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+#
+# BSD 3-Clause License
+#
+# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+#   list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+#   this list of conditions and the following disclaimer in the documentation
+#   and/or other materials provided with the distribution.
+#
+# * Neither the name of the copyright holder nor the names of its
+#   contributors may be used to endorse or promote products derived from
+#   this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+"""Inference-only Phi-1.5 model compatible with HuggingFace weights."""
+from typing import List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               LinearMethodBase,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class PhiEmbedding(nn.Module):
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+
+        self.wte = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+
+    def forward(self, input_ids: torch.LongTensor):
+        return self.wte(input_ids)
+
+
+class PhiAttention(nn.Module):
+
+    def __init__(self,
+                 config: PretrainedConfig,
+                 linear_method: Optional[LinearMethodBase] = None):
+        super().__init__()
+        self.total_num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.total_num_heads
+
+        tensor_model_parallel_world_size = (
+            get_tensor_model_parallel_world_size())
+        assert self.total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = (self.total_num_heads //
+                          tensor_model_parallel_world_size)
+
+        # pylint: disable=C0103
+        self.Wqkv = QKVParallelLinear(
+            self.hidden_size,
+            self.head_size,
+            self.total_num_heads,
+            linear_method=linear_method,
+        )
+        self.qkv_proj = QKVParallelLinear(
+            config.hidden_size,
+            self.head_size,
+            self.total_num_heads,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.out_proj = RowParallelLinear(
+            self.hidden_size,
+            self.hidden_size,
+            linear_method=linear_method,
+        )
+
+        scaling = self.head_size**-0.5
+        rotary_dim = config.rotary_dim
+        assert rotary_dim % 2 == 0
+
+        # pylint: disable=C0301
+        # Refer to:
+        # https://huggingface.co/microsoft/phi-1_5/blob/d212a789620c380ff32ca1d1ee9943a777360987/modeling_phi.py#L518
+        rope_theta = 10000
+        max_position_embeddings = getattr(config, "n_positions", 2048)
+        self.rotary_emb = get_rope(
+            self.head_size,
+            rotary_dim=rotary_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+        )
+        self.attn = PagedAttention(self.num_heads, self.head_size, scaling)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.Wqkv(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        q, k = self.rotary_emb(position_ids, q, k)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.out_proj(attn_output)
+        return output
+
+
+class PhiMLP(nn.Module):
+
+    def __init__(self,
+                 config: PretrainedConfig,
+                 linear_method: Optional[LinearMethodBase] = None):
+        super().__init__()
+
+        n_inner = getattr(config, "n_inner", None)
+        n_inner = n_inner if n_inner is not None else 4 * config.hidden_size
+
+        self.fc1 = ColumnParallelLinear(
+            config.hidden_size,
+            n_inner,
+            linear_method=linear_method,
+        )
+        self.fc2 = RowParallelLinear(
+            n_inner,
+            config.hidden_size,
+            linear_method=linear_method,
+        )
+        quant_config = getattr(linear_method, "quant_config", None)
+        self.act = get_act_fn(config.activation_function, quant_config,
+                              n_inner)
+
+    def forward(self, hidden_states):
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.fc2(hidden_states)
+        return hidden_states
+
+
+class PhiLayer(nn.Module):
+
+    def __init__(self,
+                 config: PretrainedConfig,
+                 linear_method: Optional[LinearMethodBase] = None):
+        super().__init__()
+        self.ln = nn.LayerNorm(config.hidden_size,
+                               eps=config.layer_norm_epsilon)
+        self.mixer = PhiAttention(config, linear_method)
+        self.mlp = PhiMLP(config, linear_method)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.ln(hidden_states)
+        attn_outputs = self.mixer(
+            position_ids=position_ids,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            input_metadata=input_metadata,
+        )
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        hidden_states = attn_outputs + feed_forward_hidden_states + residual
+        return hidden_states
+
+
+class PhiModel(nn.Module):
+
+    def __init__(self,
+                 config: PretrainedConfig,
+                 linear_method: Optional[LinearMethodBase] = None):
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.embd = PhiEmbedding(config)
+        self.h = nn.ModuleList([
+            PhiLayer(config, linear_method)
+            for _ in range(config.num_hidden_layers)
+        ])
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embd(input_ids)
+        for i in range(self.config.num_hidden_layers):
+            layer = self.h[i]
+            hidden_states = layer(
+                positions,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+            )
+        return hidden_states
+
+
+class PhiCausalLMHead(nn.Module):
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+        self.ln = nn.LayerNorm(config.hidden_size,
+                               eps=config.layer_norm_epsilon)
+        self.linear = ParallelLMHead(config.vocab_size,
+                                     config.hidden_size,
+                                     bias=True)
+
+
+class PhiForCausalLM(nn.Module):
+
+    def __init__(self,
+                 config: PretrainedConfig,
+                 linear_method: Optional[LinearMethodBase] = None):
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+
+        self.transformer = PhiModel(config, linear_method)
+        self.lm_head = PhiCausalLMHead(config)
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.transformer(input_ids, positions, kv_caches,
+                                         input_metadata)
+        hidden_states = self.lm_head.ln(hidden_states)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        head = self.lm_head.linear
+        next_tokens = self.sampler(head.weight, hidden_states,
+                                   sampling_metadata, head.bias)
+        return next_tokens
+
+    def load_weights(self,
+                     model_name_or_path: str,
+                     cache_dir: Optional[str] = None,
+                     load_format: str = "auto",
+                     revision: Optional[str] = None):
+        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:
+                continue
+
+            # Skip loading extra bias for GPTQ models.
+            if name.endswith(".bias") and name not in params_dict:
+                continue
+            # pylint: disable=E1136
+            param = params_dict[name]
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)

vllm/model_executor/models/qwen.py

+# coding=utf-8
+# Adapted from
+# https://huggingface.co/Qwen/Qwen-7B/blob/main/modeling_qwen.py
+# Copyright (c) Alibaba Cloud.
+# LICENSE: https://huggingface.co/Qwen/Qwen-7B/blob/main/LICENSE
+"""Inference-only QWen model compatible with HuggingFace weights."""
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from torch import nn
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.attention import PagedAttention
+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.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+from vllm.transformers_utils.configs.qwen import QWenConfig
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class QWenMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str = "silu",
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2,
+            bias=False,
+            linear_method=linear_method)
+        self.c_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.c_proj(x)
+        return x
+
+
+class QWenAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        max_position_embeddings: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
+        )
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = (self.total_num_heads //
+                          tensor_model_parallel_world_size)
+        self.head_dim = hidden_size // self.total_num_heads
+        self.c_attn = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            bias=True,
+            linear_method=linear_method,
+        )
+        self.c_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.scaling = self.head_dim**-0.5
+
+        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 = PagedAttention(self.num_heads, self.head_dim, self.scaling)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.c_attn(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+
+        output, _ = self.c_proj(attn_output)
+        return output
+
+
+class QWenBlock(nn.Module):
+
+    def __init__(
+        self,
+        config: QWenConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        self.attn = QWenAttention(config.hidden_size,
+                                  config.num_attention_heads,
+                                  config.max_position_embeddings,
+                                  rope_theta=rope_theta,
+                                  rope_scaling=rope_scaling,
+                                  linear_method=linear_method)
+
+        self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+        self.mlp = QWenMLP(config.hidden_size,
+                           config.intermediate_size // 2,
+                           linear_method=linear_method)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.ln_1(hidden_states)
+        else:
+            hidden_states, residual = self.ln_1(hidden_states, residual)
+        hidden_states = self.attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.ln_2(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class QWenModel(nn.Module):
+
+    def __init__(
+        self,
+        config: QWenConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+
+        self.wte = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.h = nn.ModuleList([
+            QWenBlock(config, linear_method)
+            for _ in range(config.num_hidden_layers)
+        ])
+        self.ln_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.wte(input_ids)
+        residual = None
+        for i in range(len(self.h)):
+            layer = self.h[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.ln_f(hidden_states, residual)
+        return hidden_states
+
+
+class QWenLMHeadModel(nn.Module):
+
+    def __init__(
+        self,
+        config: QWenConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.transformer = QWenModel(config, linear_method)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.transformer(input_ids, positions, kv_caches,
+                                         input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    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)
+            ("gate_up_proj", "w2", 0),
+            ("gate_up_proj", "w1", 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:
+                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)

vllm/model_executor/models/yi.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only Yi model (https://01.ai) compatible with HuggingFace weights."""
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from torch import nn
+from vllm.transformers_utils.configs.yi import YiConfig
+
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.attention import PagedAttention
+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.sampler import Sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding, ParallelLMHead)
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.weight_utils import (default_weight_loader,
+                                              hf_model_weights_iterator)
+from vllm.sequence import SamplerOutput
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class YiMLP(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 YiAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        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=self.rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = PagedAttention(self.num_heads,
+                                   self.head_dim,
+                                   self.scaling,
+                                   num_kv_heads=self.num_kv_heads)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        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)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class YiDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: YiConfig,
+        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 = YiAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            linear_method=linear_method,
+        )
+        self.mlp = YiMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            linear_method=linear_method,
+        )
+        self.ln1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.ln2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.ln1(hidden_states)
+        else:
+            hidden_states, residual = self.ln1(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.ln2(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class YiModel(nn.Module):
+
+    def __init__(
+        self,
+        config: YiConfig,
+        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([
+            YiDecoderLayer(config, linear_method)
+            for _ 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,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class YiForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: YiConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.model = YiModel(config, linear_method)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.sampler = Sampler(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   input_metadata)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
+        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
+                                   sampling_metadata)
+        return next_tokens
+
+    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:
+                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)

vllm/model_executor/parallel_utils/README.md

+The files in this folder are ported from [Megatron-LM](https://github.com/NVIDIA/Megatron-LM/tree/main/megatron/core). We only keep the codes that are used in inference.
\ No newline at end of file

vllm/model_executor/parallel_utils/communication_op.py

+import torch
+
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size,
+    get_tensor_model_parallel_group,
+)
+
+
+def tensor_model_parallel_all_reduce(input_):
+    """All-reduce the input tensor across model parallel group.
+
+    NOTE: This operation is applied in-place on the input tensor.
+    """
+    # Bypass the function if we are using only 1 GPU.
+    if get_tensor_model_parallel_world_size() == 1:
+        return input_
+    # All-reduce.
+    torch.distributed.all_reduce(input_,
+                                 group=get_tensor_model_parallel_group())
+    return input_
+
+
+def tensor_model_parallel_all_gather(input_, dim=-1):
+    """All-gather the input tensor across model parallel group."""
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size == 1:
+        return input_
+    assert -input_.dim() <= dim < input_.dim(), (
+        f"Invalid dim ({dim}) for input tensor with shape {input_.size()}")
+    if dim < 0:
+        # Convert negative dim to positive.
+        dim += input_.dim()
+    input_size = input_.size()
+    # Allocate output tensor.
+    output_tensor = torch.empty((world_size, ) + input_size,
+                                dtype=input_.dtype,
+                                device=input_.device)
+    # All-gather.
+    torch.distributed.all_gather_into_tensor(
+        output_tensor, input_, group=get_tensor_model_parallel_group())
+    # Reshape
+    output_tensor = output_tensor.movedim(0, dim)
+    output_tensor = output_tensor.reshape(input_size[:dim] +
+                                          (world_size * input_size[dim], ) +
+                                          input_size[dim + 1:])
+    return output_tensor

vllm/model_executor/parallel_utils/parallel_state.py

+# Copyright 2023 The vLLM team.
+# Adapted from
+# https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/parallel_state.py
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+"""Tensor and pipeline parallel groups."""
+
+import torch
+
+# Tensor model parallel group that the current rank belongs to.
+_TENSOR_MODEL_PARALLEL_GROUP = None
+# Pipeline model parallel group that the current rank belongs to.
+_PIPELINE_MODEL_PARALLEL_GROUP = None
+
+# A list of global ranks for each pipeline group to ease calculation of the
+# source rank when broadcasting from the first or last pipeline stage.
+_PIPELINE_GLOBAL_RANKS = None
+
+
+def initialize_model_parallel(
+    tensor_model_parallel_size: int = 1,
+    pipeline_model_parallel_size: int = 1,
+) -> None:
+    """
+    Initialize model parallel groups.
+
+    Arguments:
+        tensor_model_parallel_size: number of GPUs used for tensor model
+            parallelism.
+        pipeline_model_parallel_size: number of GPUs used for pipeline model
+            parallelism.
+
+    Let's say we have a total of 8 GPUs denoted by g0 ... g7 and we
+    use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
+    the model pipeline. The present function will
+    create 4 tensor model-parallel groups and 2 pipeline model-parallel groups:
+        4 tensor model-parallel groups:
+            [g0, g1], [g2, g3], [g4, g5], [g6, g7]
+        2 pipeline model-parallel groups:
+            [g0, g2, g4, g6], [g1, g3, g5, g7]
+    Note that for efficiency, the caller should make sure adjacent ranks
+    are on the same DGX box. For example if we are using 2 DGX-1 boxes
+    with a total of 16 GPUs, rank 0 to 7 belong to the first box and
+    ranks 8 to 15 belong to the second box.
+    """
+    # Get world size and rank. Ensure some consistencies.
+    assert torch.distributed.is_initialized()
+    world_size: int = torch.distributed.get_world_size()
+
+    if (world_size !=
+            tensor_model_parallel_size * pipeline_model_parallel_size):
+        raise RuntimeError(
+            f"world_size ({world_size}) is not equal to "
+            f"tensor_model_parallel_size ({tensor_model_parallel_size}) x "
+            f"pipeline_model_parallel_size ({pipeline_model_parallel_size})")
+
+    num_tensor_model_parallel_groups: int = (world_size //
+                                             tensor_model_parallel_size)
+    num_pipeline_model_parallel_groups: int = (world_size //
+                                               pipeline_model_parallel_size)
+    rank = torch.distributed.get_rank()
+
+    # Build the tensor model-parallel groups.
+    global _TENSOR_MODEL_PARALLEL_GROUP
+    assert _TENSOR_MODEL_PARALLEL_GROUP is None, (
+        "tensor model parallel group is already initialized")
+    for i in range(num_tensor_model_parallel_groups):
+        ranks = range(i * tensor_model_parallel_size,
+                      (i + 1) * tensor_model_parallel_size)
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _TENSOR_MODEL_PARALLEL_GROUP = group
+
+    # Build the pipeline model-parallel groups.
+    global _PIPELINE_MODEL_PARALLEL_GROUP
+    global _PIPELINE_GLOBAL_RANKS
+    assert _PIPELINE_MODEL_PARALLEL_GROUP is None, (
+        "pipeline model parallel group is already initialized")
+    for i in range(num_pipeline_model_parallel_groups):
+        ranks = range(i, world_size, num_pipeline_model_parallel_groups)
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _PIPELINE_MODEL_PARALLEL_GROUP = group
+            _PIPELINE_GLOBAL_RANKS = ranks
+
+
+def model_parallel_is_initialized():
+    """Check if tensor and pipeline parallel groups are initialized."""
+    return (_TENSOR_MODEL_PARALLEL_GROUP is not None
+            and _PIPELINE_MODEL_PARALLEL_GROUP is not None)
+
+
+def get_tensor_model_parallel_group():
+    """Get the tensor model parallel group the caller rank belongs to."""
+    assert _TENSOR_MODEL_PARALLEL_GROUP is not None, (
+        "tenosr model parallel group is not initialized")
+    return _TENSOR_MODEL_PARALLEL_GROUP
+
+
+def get_pipeline_model_parallel_group():
+    """Get the pipeline model parallel group the caller rank belongs to."""
+    assert _PIPELINE_MODEL_PARALLEL_GROUP is not None, (
+        "pipeline model parallel group is not initialized")
+    return _PIPELINE_MODEL_PARALLEL_GROUP
+
+
+def get_tensor_model_parallel_world_size():
+    """Return world size for the tensor model parallel group."""
+    return torch.distributed.get_world_size(
+        group=get_tensor_model_parallel_group())
+
+
+def get_pipeline_model_parallel_world_size():
+    """Return world size for the pipeline model parallel group."""
+    return torch.distributed.get_world_size(
+        group=get_pipeline_model_parallel_group())
+
+
+def get_tensor_model_parallel_rank():
+    """Return my rank for the tensor model parallel group."""
+    return torch.distributed.get_rank(group=get_tensor_model_parallel_group())
+
+
+def get_pipeline_model_parallel_rank():
+    """Return my rank for the pipeline model parallel group."""
+    return torch.distributed.get_rank(
+        group=get_pipeline_model_parallel_group())
+
+
+def get_tensor_model_parallel_src_rank():
+    """Calculate the global rank corresponding to the first local rank
+    in the tensor model parallel group."""
+    global_rank = torch.distributed.get_rank()
+    local_world_size = get_tensor_model_parallel_world_size()
+    return (global_rank // local_world_size) * local_world_size
+
+
+def get_pipeline_model_parallel_first_rank():
+    """Return the global rank of the first process in the pipeline for the
+    current tensor parallel group"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, (
+        "Pipeline parallel group is not initialized")
+    return _PIPELINE_GLOBAL_RANKS[0]
+
+
+def get_pipeline_model_parallel_last_rank():
+    """Return the global rank of the last process in the pipeline for the
+    current tensor parallel group"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, (
+        "Pipeline parallel group is not initialized")
+    last_rank_local = get_pipeline_model_parallel_world_size() - 1
+    return _PIPELINE_GLOBAL_RANKS[last_rank_local]
+
+
+def get_pipeline_model_parallel_next_rank():
+    """Return the global rank that follows the caller in the pipeline"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, (
+        "Pipeline parallel group is not initialized")
+    rank_in_pipeline = get_pipeline_model_parallel_rank()
+    world_size = get_pipeline_model_parallel_world_size()
+    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline + 1) % world_size]
+
+
+def get_pipeline_model_parallel_prev_rank():
+    """Return the global rank that preceeds the caller in the pipeline"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, (
+        "Pipeline parallel group is not initialized")
+    rank_in_pipeline = get_pipeline_model_parallel_rank()
+    world_size = get_pipeline_model_parallel_world_size()
+    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline - 1) % world_size]
+
+
+def destroy_model_parallel():
+    """Set the groups to none."""
+    global _TENSOR_MODEL_PARALLEL_GROUP
+    _TENSOR_MODEL_PARALLEL_GROUP = None
+    global _PIPELINE_MODEL_PARALLEL_GROUP
+    _PIPELINE_MODEL_PARALLEL_GROUP = None
+    global _PIPELINE_GLOBAL_RANKS
+    _PIPELINE_GLOBAL_RANKS = None

vllm/model_executor/parallel_utils/utils.py

+# Copyright 2023 The vLLM team.
+# Adapted from
+# https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/tensor_parallel/utils.py
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+from typing import Sequence
+
+import torch
+
+
+def ensure_divisibility(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator."""
+    assert numerator % denominator == 0, "{} is not divisible by {}".format(
+        numerator, denominator)
+
+
+def divide(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator and return
+    the division value."""
+    ensure_divisibility(numerator, denominator)
+    return numerator // denominator
+
+
+def split_tensor_along_last_dim(
+    tensor: torch.Tensor,
+    num_partitions: int,
+    contiguous_split_chunks: bool = False,
+) -> Sequence[torch.Tensor]:
+    """ Split a tensor along its last dimension.
+
+        Arguments:
+            tensor: input tensor.
+            num_partitions: number of partitions to split the tensor
+            contiguous_split_chunks: If True, make each chunk contiguous
+                                     in memory.
+
+        Returns:
+            A list of Tensors
+    """
+    # Get the size and dimension.
+    last_dim = tensor.dim() - 1
+    last_dim_size = divide(tensor.size()[last_dim], num_partitions)
+    # Split.
+    tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+    # NOTE: torch.split does not create contiguous tensors by default.
+    if contiguous_split_chunks:
+        return tuple(chunk.contiguous() for chunk in tensor_list)
+
+    return tensor_list

vllm/model_executor/sampling_metadata.py

+from dataclasses import dataclass
+from typing import Dict, List, Tuple
+
+import torch
+
+from vllm.sampling_params import SamplingParams, SamplingType
+from vllm.sequence import SequenceData
+from vllm.utils import in_wsl
+
+_SAMPLING_EPS = 1e-5
+
+
+class SamplingMetadata:
+    """Metadata for input sequences. Used in sampler.
+
+    Args:
+        seq_groups: List of (seq_ids, sampling_params).
+        seq_data: Seq_id -> SequenceData.
+        prompt_lens: Lengths of prompts.
+        selected_token_indices: Token indices selected for sampling.
+        categorized_sample_indices: SamplingType -> token indicies to sample.
+    """
+
+    def __init__(
+        self,
+        seq_groups: List[Tuple[List[int], SamplingParams]],
+        seq_data: Dict[int, SequenceData],
+        prompt_lens: List[int],
+        selected_token_indices: torch.Tensor,
+        categorized_sample_indices: Dict[SamplingType, torch.Tensor],
+    ) -> None:
+        self.seq_groups = seq_groups
+        self.seq_data = seq_data
+        self.prompt_lens = prompt_lens
+        self.selected_token_indices = selected_token_indices
+        self.categorized_sample_indices = categorized_sample_indices
+
+        self.num_prompts = len(prompt_lens)
+
+    def __repr__(self) -> str:
+        return (
+            "SamplingMetadata("
+            f"seq_groups={self.seq_groups}, "
+            f"seq_data={self.seq_data}, "
+            f"prompt_lens={self.prompt_lens}, "
+            f"selected_token_indices={self.selected_token_indices}, "
+            f"categorized_sample_indices={self.categorized_sample_indices})")
+
+
+@dataclass
+class SamplingTensors:
+    """Tensors for sampling."""
+
+    temperatures: torch.Tensor
+    top_ps: torch.Tensor
+    top_ks: torch.Tensor
+    min_ps: torch.Tensor
+    presence_penalties: torch.Tensor
+    frequency_penalties: torch.Tensor
+    repetition_penalties: torch.Tensor
+    prompt_tokens: torch.Tensor
+    output_tokens: torch.Tensor
+
+    @classmethod
+    def from_sampling_metadata(
+            cls, sampling_metadata: "SamplingMetadata", vocab_size: int,
+            device: torch.device,
+            dtype: torch.dtype) -> Tuple["SamplingTensors", bool, bool, bool]:
+        prompt_tokens: List[List[int]] = []
+        output_tokens: List[List[int]] = []
+        top_ks: List[int] = []
+        temperatures: List[float] = []
+        top_ps: List[float] = []
+        min_ps: List[float] = []
+        presence_penalties: List[float] = []
+        frequency_penalties: List[float] = []
+        repetition_penalties: List[float] = []
+        do_penalties = False
+        do_top_p_top_k = False
+        do_min_p = False
+        for i, seq_group in enumerate(sampling_metadata.seq_groups):
+            seq_ids, sampling_params = seq_group
+            temperature = sampling_params.temperature
+            p = sampling_params.presence_penalty
+            f = sampling_params.frequency_penalty
+            r = sampling_params.repetition_penalty
+            top_p = sampling_params.top_p
+            min_p = sampling_params.min_p
+            # k should not be greater than the vocab size.
+            top_k = min(sampling_params.top_k, vocab_size)
+            top_k = vocab_size if top_k == -1 else top_k
+            if temperature < _SAMPLING_EPS:
+                # NOTE: Zero temperature means deterministic sampling
+                # (i.e., greedy sampling or beam search).
+                # Set the temperature to 1 to avoid division by zero.
+                temperature = 1.0
+            if not do_top_p_top_k and (top_p < 1.0 - _SAMPLING_EPS
+                                       or top_k != vocab_size):
+                do_top_p_top_k = True
+            if not do_min_p and min_p > _SAMPLING_EPS:
+                do_min_p = True
+            if not do_penalties and (abs(p) >= _SAMPLING_EPS
+                                     or abs(f) >= _SAMPLING_EPS
+                                     or abs(r - 1.0) >= _SAMPLING_EPS):
+                do_penalties = True
+            if (i < sampling_metadata.num_prompts
+                    and sampling_params.prompt_logprobs is not None):
+                # For tokens in the prompt that we only need to get their logprobs
+                prompt_len = sampling_metadata.prompt_lens[i]
+                temperatures += [temperature] * (prompt_len - 1)
+                top_ps += [top_p] * (prompt_len - 1)
+                top_ks += [top_k] * (prompt_len - 1)
+                min_ps += [min_p] * (prompt_len - 1)
+                presence_penalties += [0] * (prompt_len - 1)
+                frequency_penalties += [0] * (prompt_len - 1)
+                repetition_penalties += [1] * (prompt_len - 1)
+                prompt_tokens.extend([] for _ in range(prompt_len - 1))
+                output_tokens.extend([] for _ in range(prompt_len - 1))
+            for seq_id in seq_ids:
+                seq_data = sampling_metadata.seq_data[seq_id]
+                prompt_tokens.append(seq_data.prompt_token_ids)
+                output_tokens.append(seq_data.output_token_ids)
+            temperatures += [temperature] * len(seq_ids)
+            top_ps += [top_p] * len(seq_ids)
+            top_ks += [top_k] * len(seq_ids)
+            min_ps += [min_p] * len(seq_ids)
+            presence_penalties += [p] * len(seq_ids)
+            frequency_penalties += [f] * len(seq_ids)
+            repetition_penalties += [r] * len(seq_ids)
+
+        sampling_tensors = SamplingTensors.from_lists(
+            temperatures, top_ps, top_ks, min_ps, presence_penalties,
+            frequency_penalties, repetition_penalties, prompt_tokens,
+            output_tokens, vocab_size, device, dtype)
+        return (sampling_tensors, do_penalties, do_top_p_top_k, do_min_p)
+
+    @classmethod
+    def from_lists(cls, temperatures: List[float], top_ps: List[float],
+                   top_ks: List[int], min_ps: List[float],
+                   presence_penalties: List[float],
+                   frequency_penalties: List[float],
+                   repetition_penalties: List[float],
+                   prompt_tokens: List[List[int]],
+                   output_tokens: List[List[int]], vocab_size: int,
+                   device: torch.device,
+                   dtype: torch.dtype) -> "SamplingTensors":
+        # Note that the performance will be very bad without
+        # pinned memory.
+        pin_memory = not in_wsl()
+        prompt_max_len = max(len(tokens) for tokens in prompt_tokens)
+        prompt_padded_tokens = [
+            tokens + [vocab_size] * (prompt_max_len - len(tokens))
+            for tokens in prompt_tokens
+        ]
+        output_max_len = max(len(tokens) for tokens in output_tokens)
+        output_padded_tokens = [
+            tokens + [vocab_size] * (output_max_len - len(tokens))
+            for tokens in output_tokens
+        ]
+
+        temperatures_t = torch.tensor(
+            temperatures,
+            device="cpu",
+            dtype=dtype,
+            pin_memory=pin_memory,
+        )
+        top_ps_t = torch.tensor(
+            top_ps,
+            device="cpu",
+            dtype=dtype,
+            pin_memory=pin_memory,
+        )
+        min_ps_t = torch.tensor(
+            min_ps,
+            device="cpu",
+            dtype=dtype,
+            pin_memory=pin_memory,
+        )
+        presence_penalties_t = torch.tensor(
+            presence_penalties,
+            device="cpu",
+            dtype=dtype,
+            pin_memory=pin_memory,
+        )
+        frequency_penalties_t = torch.tensor(
+            frequency_penalties,
+            device="cpu",
+            dtype=dtype,
+            pin_memory=pin_memory,
+        )
+        repetition_penalties_t = torch.tensor(
+            repetition_penalties,
+            device="cpu",
+            dtype=dtype,
+            pin_memory=pin_memory,
+        )
+        top_ks_t = torch.tensor(
+            top_ks,
+            device="cpu",
+            dtype=torch.int,
+            pin_memory=pin_memory,
+        )
+        prompt_tensor = torch.tensor(
+            prompt_padded_tokens,
+            device="cpu",
+            dtype=torch.long,
+            pin_memory=pin_memory,
+        )
+        output_tensor = torch.tensor(
+            output_padded_tokens,
+            device="cpu",
+            dtype=torch.long,
+            pin_memory=pin_memory,
+        )
+        # Because the memory is pinned, we can do non-blocking
+        # transfer to device.
+        return cls(
+            temperatures=temperatures_t.to(device=device, non_blocking=True),
+            top_ps=top_ps_t.to(device=device, non_blocking=True),
+            top_ks=top_ks_t.to(device=device, non_blocking=True),
+            min_ps=min_ps_t.to(device=device, non_blocking=True),
+            presence_penalties=presence_penalties_t.to(device=device,
+                                                       non_blocking=True),
+            frequency_penalties=frequency_penalties_t.to(device=device,
+                                                         non_blocking=True),
+            repetition_penalties=repetition_penalties_t.to(device=device,
+                                                           non_blocking=True),
+            prompt_tokens=prompt_tensor.to(device=device, non_blocking=True),
+            output_tokens=output_tensor.to(device=device, non_blocking=True),
+        )

vllm/model_executor/utils.py

+"""Utils for model executor."""
+import random
+from typing import Any, Dict, Optional
+
+import numpy as np
+import torch
+
+
+def set_random_seed(seed: int) -> None:
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+
+def set_weight_attrs(
+    weight: torch.Tensor,
+    weight_attrs: Optional[Dict[str, Any]],
+):
+    """Set attributes on a weight tensor.
+
+    This method is used to set attributes on a weight tensor. This method
+    will not overwrite existing attributes.
+
+    Args:
+        weight: The weight tensor.
+        weight_attrs: A dictionary of attributes to set on the weight tensor.
+    """
+    if weight_attrs is None:
+        return
+    for key, value in weight_attrs.items():
+        assert not hasattr(
+            weight, key), (f"Overwriting existing tensor attribute: {key}")
+        setattr(weight, key, value)

vllm/model_executor/weight_utils.py

+"""Utilities for downloading and initializing model weights."""
+import filelock
+import glob
+import json
+import os
+from collections import defaultdict
+from typing import Any, Iterator, List, Optional, Tuple
+
+from huggingface_hub import snapshot_download
+import numpy as np
+from safetensors.torch import load_file, save_file, safe_open
+import torch
+from transformers import PretrainedConfig
+from tqdm.auto import tqdm
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization import (get_quantization_config,
+                                                     QuantizationConfig)
+
+logger = init_logger(__name__)
+
+
+class Disabledtqdm(tqdm):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs, disable=True)
+
+
+def get_lock(model_name_or_path: str, cache_dir: Optional[str] = None):
+    lock_dir = cache_dir if cache_dir is not None else "/tmp"
+    lock_file_name = model_name_or_path.replace("/", "-") + ".lock"
+    lock = filelock.FileLock(os.path.join(lock_dir, lock_file_name))
+    return lock
+
+
+def _shared_pointers(tensors):
+    ptrs = defaultdict(list)
+    for k, v in tensors.items():
+        ptrs[v.data_ptr()].append(k)
+    failing = []
+    for _, names in ptrs.items():
+        if len(names) > 1:
+            failing.append(names)
+    return failing
+
+
+def convert_bin_to_safetensor_file(
+    pt_filename: str,
+    sf_filename: str,
+) -> None:
+    loaded = torch.load(pt_filename, map_location="cpu")
+    if "state_dict" in loaded:
+        loaded = loaded["state_dict"]
+    shared = _shared_pointers(loaded)
+    for shared_weights in shared:
+        for name in shared_weights[1:]:
+            loaded.pop(name)
+
+    # For tensors to be contiguous
+    loaded = {k: v.contiguous() for k, v in loaded.items()}
+
+    dirname = os.path.dirname(sf_filename)
+    os.makedirs(dirname, exist_ok=True)
+    save_file(loaded, sf_filename, metadata={"format": "pt"})
+
+    # check file size
+    sf_size = os.stat(sf_filename).st_size
+    pt_size = os.stat(pt_filename).st_size
+    if (sf_size - pt_size) / pt_size > 0.01:
+        raise RuntimeError(f"""The file size different is more than 1%:
+         - {sf_filename}: {sf_size}
+         - {pt_filename}: {pt_size}
+         """)
+
+    # check if the tensors are the same
+    reloaded = load_file(sf_filename)
+    for k in loaded:
+        pt_tensor = loaded[k]
+        sf_tensor = reloaded[k]
+        if not torch.equal(pt_tensor, sf_tensor):
+            raise RuntimeError(f"The output tensors do not match for key {k}")
+
+
+# TODO(woosuk): Move this to other place.
+def get_quant_config(
+    quantization: str,
+    model_name_or_path: str,
+    hf_config: PretrainedConfig,
+    cache_dir: Optional[str] = None,
+) -> QuantizationConfig:
+    quant_cls = get_quantization_config(quantization)
+    # Read the quantization config from the HF model config, if available.
+    hf_quant_config = getattr(hf_config, "quantization_config", None)
+    if hf_quant_config is not None:
+        return quant_cls.from_config(hf_quant_config)
+
+    is_local = os.path.isdir(model_name_or_path)
+    if not is_local:
+        # Download the config files.
+        with get_lock(model_name_or_path, cache_dir):
+            hf_folder = snapshot_download(model_name_or_path,
+                                          allow_patterns="*.json",
+                                          cache_dir=cache_dir,
+                                          tqdm_class=Disabledtqdm)
+    else:
+        hf_folder = model_name_or_path
+    config_files = glob.glob(os.path.join(hf_folder, "*.json"))
+
+    quant_config_files = [
+        f for f in config_files if any(
+            f.endswith(x) for x in quant_cls.get_config_filenames())
+    ]
+    if len(quant_config_files) == 0:
+        raise ValueError(f"Cannot find the config file for {quantization}")
+    if len(quant_config_files) > 1:
+        raise ValueError(f"Found multiple config files for {quantization}: "
+                         f"{quant_config_files}")
+
+    quant_config_file = quant_config_files[0]
+    with open(quant_config_file, "r") as f:
+        config = json.load(f)
+    return quant_cls.from_config(config)
+
+
+def prepare_hf_model_weights(
+    model_name_or_path: str,
+    cache_dir: Optional[str] = None,
+    load_format: str = "auto",
+    fall_back_to_pt: bool = True,
+    revision: Optional[str] = None,
+) -> Tuple[str, List[str], bool]:
+    # Download model weights from huggingface.
+    is_local = os.path.isdir(model_name_or_path)
+    use_safetensors = False
+    # Some quantized models use .pt files for storing the weights.
+    if load_format == "auto":
+        allow_patterns = ["*.safetensors", "*.bin"]
+    elif load_format == "safetensors":
+        use_safetensors = True
+        allow_patterns = ["*.safetensors"]
+    elif load_format == "pt":
+        allow_patterns = ["*.pt"]
+    elif load_format == "npcache":
+        allow_patterns = ["*.bin"]
+    else:
+        raise ValueError(f"Unknown load_format: {load_format}")
+
+    if fall_back_to_pt:
+        allow_patterns += ["*.pt"]
+
+    if not is_local:
+        # Use file lock to prevent multiple processes from
+        # downloading the same model weights at the same time.
+        with get_lock(model_name_or_path, cache_dir):
+            hf_folder = snapshot_download(model_name_or_path,
+                                          allow_patterns=allow_patterns,
+                                          cache_dir=cache_dir,
+                                          tqdm_class=Disabledtqdm,
+                                          revision=revision)
+    else:
+        hf_folder = model_name_or_path
+    hf_weights_files: List[str] = []
+    for pattern in allow_patterns:
+        hf_weights_files += glob.glob(os.path.join(hf_folder, pattern))
+        if len(hf_weights_files) > 0:
+            if pattern == "*.safetensors":
+                use_safetensors = True
+            break
+    if not use_safetensors:
+        # Exclude files that are not needed for inference.
+        # https://github.com/huggingface/transformers/blob/v4.34.0/src/transformers/trainer.py#L227-L233
+        blacklist = [
+            "training_args.bin",
+            "optimizer.bin",
+            "optimizer.pt",
+            "scheduler.pt",
+            "scaler.pt",
+        ]
+        hf_weights_files = [
+            f for f in hf_weights_files
+            if not any(f.endswith(x) for x in blacklist)
+        ]
+
+    if len(hf_weights_files) == 0:
+        raise RuntimeError(
+            f"Cannot find any model weights with `{model_name_or_path}`")
+
+    return hf_folder, hf_weights_files, use_safetensors
+
+
+def hf_model_weights_iterator(
+    model_name_or_path: str,
+    cache_dir: Optional[str] = None,
+    load_format: str = "auto",
+    revision: Optional[str] = None,
+    fall_back_to_pt: Optional[bool] = True,
+) -> Iterator[Tuple[str, torch.Tensor]]:
+    hf_folder, hf_weights_files, use_safetensors = prepare_hf_model_weights(
+        model_name_or_path,
+        cache_dir=cache_dir,
+        load_format=load_format,
+        fall_back_to_pt=fall_back_to_pt,
+        revision=revision)
+
+    if load_format == "npcache":
+        # Currently np_cache only support *.bin checkpoints
+        assert use_safetensors is False
+
+        # Convert the model weights from torch tensors to numpy arrays for
+        # faster loading.
+        np_folder = os.path.join(hf_folder, "np")
+        os.makedirs(np_folder, exist_ok=True)
+        weight_names_file = os.path.join(np_folder, "weight_names.json")
+        # Use file lock to prevent multiple processes from
+        # dumping the same model weights to numpy at the same time.
+        with get_lock(model_name_or_path, cache_dir):
+            if not os.path.exists(weight_names_file):
+                weight_names = []
+                for bin_file in hf_weights_files:
+                    state = torch.load(bin_file, map_location="cpu")
+                    for name, param in state.items():
+                        param_path = os.path.join(np_folder, name)
+                        with open(param_path, "wb") as f:
+                            np.save(f, param.cpu().detach().numpy())
+                        weight_names.append(name)
+                with open(weight_names_file, "w") as f:
+                    json.dump(weight_names, f)
+
+        with open(weight_names_file, "r") as f:
+            weight_names = json.load(f)
+
+        for name in weight_names:
+            param_path = os.path.join(np_folder, name)
+            with open(param_path, "rb") as f:
+                param = np.load(f)
+            yield name, torch.from_numpy(param)
+    elif use_safetensors:
+        for st_file in hf_weights_files:
+            with safe_open(st_file, framework="pt") as f:
+                for name in f.keys():  # noqa: SIM118
+                    param = f.get_tensor(name)
+                    yield name, param
+    else:
+        for bin_file in hf_weights_files:
+            state = torch.load(bin_file, map_location="cpu")
+            for name, param in state.items():
+                yield name, param
+            del state
+            torch.cuda.empty_cache()
+
+
+def convert_pyslice_to_tensor(x: Any) -> torch.Tensor:
+    """convert PySafeSlice object from safetensors to torch.Tensor
+
+    PySafeSlice object supports indexing, which is done before loading the
+    actual tensor and can reduce the amount of memory being read into the
+    memory. However, it does not support more advanced functionalities
+    like `.view()` or `.t()`. Therefore, if we need to modify the loaded
+    tensor with these more complicated operators, we need to convert to
+    tensor first.
+    """
+    if not isinstance(x, torch.Tensor):
+        x = x[:]
+    return x
+
+
+def default_weight_loader(param: torch.Tensor,
+                          loaded_weight: torch.Tensor) -> None:
+    """Default weight loader."""
+    assert param.size() == loaded_weight.size()
+    param.data.copy_(loaded_weight)
+
+
+def initialize_dummy_weights(
+    model: torch.nn.Module,
+    low: float = -1e-3,
+    high: float = 1e-3,
+) -> None:
+    """Initialize model weights with random values.
+
+    The model weights must be randomly initialized for accurate performance
+    measurements. Additionally, the model weights should not cause NaNs in the
+    forward pass. We empirically found that initializing the weights with
+    values between -1e-3 and 1e-3 works well for most models.
+    """
+    for param in model.state_dict().values():
+        if torch.is_floating_point(param):
+            param.data.uniform_(low, high)

vllm/outputs.py

+from typing import List, Optional
+
+from vllm.sequence import (PromptLogprobs, SampleLogprobs, SequenceGroup,
+                           SequenceStatus)
+
+
+class CompletionOutput:
+    """The output data of one completion output of a request.
+
+    Args:
+        index: The index of the output in the request.
+        text: The generated output text.
+        token_ids: The token IDs of the generated output text.
+        cumulative_logprob: The cumulative log probability of the generated
+            output text.
+        logprobs: The log probabilities of the top probability words at each
+            position if the logprobs are requested.
+        finish_reason: The reason why the sequence is finished.
+    """
+
+    def __init__(
+        self,
+        index: int,
+        text: str,
+        token_ids: List[int],
+        cumulative_logprob: float,
+        logprobs: Optional[SampleLogprobs],
+        finish_reason: Optional[str] = None,
+    ) -> None:
+        self.index = index
+        self.text = text
+        self.token_ids = token_ids
+        self.cumulative_logprob = cumulative_logprob
+        self.logprobs = logprobs
+        self.finish_reason = finish_reason
+
+    def finished(self) -> bool:
+        return self.finish_reason is not None
+
+    def __repr__(self) -> str:
+        return (f"CompletionOutput(index={self.index}, "
+                f"text={self.text!r}, "
+                f"token_ids={self.token_ids}, "
+                f"cumulative_logprob={self.cumulative_logprob}, "
+                f"logprobs={self.logprobs}, "
+                f"finish_reason={self.finish_reason})")
+
+
+class RequestOutput:
+    """The output data of a request to the LLM.
+
+    Args:
+        request_id: The unique ID of the request.
+        prompt: The prompt string of the request.
+        prompt_token_ids: The token IDs of the prompt.
+        prompt_logprobs: The log probabilities to return per prompt token.
+        outputs: The output sequences of the request.
+        finished: Whether the whole request is finished.
+    """
+
+    def __init__(
+        self,
+        request_id: str,
+        prompt: str,
+        prompt_token_ids: List[int],
+        prompt_logprobs: Optional[PromptLogprobs],
+        outputs: List[CompletionOutput],
+        finished: bool,
+    ) -> None:
+        self.request_id = request_id
+        self.prompt = prompt
+        self.prompt_token_ids = prompt_token_ids
+        self.prompt_logprobs = prompt_logprobs
+        self.outputs = outputs
+        self.finished = finished
+
+    @classmethod
+    def from_seq_group(cls, seq_group: SequenceGroup) -> "RequestOutput":
+        # Get the top-n sequences.
+        n = seq_group.sampling_params.n
+        seqs = seq_group.get_seqs()
+        if seq_group.sampling_params.use_beam_search:
+            sorting_key = lambda seq: seq.get_beam_search_score(
+                seq_group.sampling_params.length_penalty)
+        else:
+            sorting_key = lambda seq: seq.get_cumulative_logprob()
+        sorted_seqs = sorted(seqs, key=sorting_key, reverse=True)
+        top_n_seqs = sorted_seqs[:n]
+
+        # Create the outputs.
+        outputs: List[CompletionOutput] = []
+        for seq in top_n_seqs:
+            logprobs = seq.output_logprobs
+            if seq_group.sampling_params.logprobs is None:
+                # NOTE: We need to take care of this case because the sequence
+                # always has the logprobs of the sampled tokens even if the
+                # logprobs are not requested.
+                logprobs = None
+            finshed_reason = SequenceStatus.get_finished_reason(seq.status)
+            output = CompletionOutput(seqs.index(seq), seq.output_text,
+                                      seq.get_output_token_ids(),
+                                      seq.get_cumulative_logprob(), logprobs,
+                                      finshed_reason)
+            outputs.append(output)
+
+        # Every sequence in the sequence group should have the same prompt.
+        prompt = seq_group.prompt
+        prompt_token_ids = seq_group.prompt_token_ids
+        prompt_logprobs = seq_group.prompt_logprobs
+        finished = seq_group.is_finished()
+        return cls(seq_group.request_id, prompt, prompt_token_ids,
+                   prompt_logprobs, outputs, finished)
+
+    def __repr__(self) -> str:
+        return (f"RequestOutput(request_id={self.request_id}, "
+                f"prompt={self.prompt!r}, "
+                f"prompt_token_ids={self.prompt_token_ids}, "
+                f"prompt_logprobs={self.prompt_logprobs}, "
+                f"outputs={self.outputs}, "
+                f"finished={self.finished})")