Merge tag 'v0.6.0' into v0.6.0-dev

vllm/model_executor/models/dbrx.py

@@ -17,13 +17,13 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs.dbrx import DbrxConfig
 
 

vllm/model_executor/models/deepseek.py

@@ -43,12 +43,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 
 class DeepseekMLP(nn.Module):

vllm/model_executor/models/deepseek_v2.py

@@ -43,12 +43,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 from .utils import PPMissingLayer, is_pp_missing_parameter, make_layers
 

vllm/model_executor/models/eagle.py

@@ -5,12 +5,13 @@ import torch.nn as nn
 
 from vllm.attention.backends.abstract import AttentionMetadata
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models import ModelRegistry
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs.eagle import EAGLEConfig
 
 

vllm/model_executor/models/exaone.py

+# coding=utf-8
+# Adapted from
+# https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/blob/main/modeling_exaone.py
+# Copyright 2024 The LG U+ CTO AI Tech Lab.
+# Copyright 2021 The LG AI Research EXAONE Lab
+# Copyright (c) 2018, NVIDIA CORPORATION.  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 Exaone model compatible with HuggingFace weights."""
+
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.config import CacheConfig, LoRAConfig
+from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
+    get_compressed_tensors_cache_scale)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader, kv_cache_scales_loader, maybe_remap_kv_scale_name)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.configs.exaone import ExaoneConfig
+from vllm.utils import is_hip
+
+from .interfaces import SupportsLoRA
+from .utils import PPMissingLayer, is_pp_missing_parameter, make_layers
+
+
+class ExaoneGatedMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            input_size=hidden_size,
+            output_sizes=[intermediate_size] * 2,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj",
+        )
+        self.c_proj = RowParallelLinear(
+            input_size=intermediate_size,
+            output_size=hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
+        )
+        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 ExaoneAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        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,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        cache_config: Optional[CacheConfig] = None,
+        prefix: str = "",
+    ) -> 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)
+        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+        self.head_dim = getattr(config, "head_dim",
+                                self.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=hidden_size,
+            head_size=self.head_dim,
+            total_num_heads=self.total_num_heads,
+            total_num_kv_heads=self.total_num_kv_heads,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+
+        self.out_proj = RowParallelLinear(
+            input_size=self.total_num_heads * self.head_dim,
+            output_size=hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.out_proj",
+        )
+
+        is_neox_style = True
+        if quant_config is not None and quant_config.get_name() == "gguf":
+            is_neox_style = False
+
+        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,
+            is_neox_style=is_neox_style,
+        )
+        self.attn = Attention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            cache_config=cache_config,
+            quant_config=quant_config,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.out_proj(attn_output)
+        return output
+
+
+class ExaoneBlockAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        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,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        cache_config: Optional[CacheConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.attention = ExaoneAttention(
+            config=config,
+            hidden_size=hidden_size,
+            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            bias=bias,
+            cache_config=cache_config,
+            prefix=prefix,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        return self.attention(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+
+class ExaoneDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+                config, "original_max_position_embeddings", None):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        # Support abacusai/Smaug-72B-v0.1 with attention_bias
+        # Support internlm/internlm-7b with bias
+        attention_bias = getattr(config, "attention_bias", False) or getattr(
+            config, "bias", False)
+        self.attn = ExaoneBlockAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=getattr(config, "num_key_value_heads",
+                                 config.num_attention_heads),
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            bias=attention_bias,
+            cache_config=cache_config,
+            prefix=f"{prefix}.attn",
+        )
+        self.mlp = ExaoneGatedMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.activation_function,
+            quant_config=quant_config,
+            bias=getattr(config, "mlp_bias", False),
+            prefix=f"{prefix}.mlp",
+        )
+        self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        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,
+            attn_metadata=attn_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.ln_2(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class ExaoneModel(nn.Module):
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        lora_config: Optional[LoRAConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        lora_vocab = ((lora_config.lora_extra_vocab_size *
+                       (lora_config.max_loras or 1)) if lora_config else 0)
+        self.vocab_size = config.vocab_size + lora_vocab
+        self.wte = config.vocab_size
+        if get_pp_group().is_first_rank or (config.tie_word_embeddings
+                                            and get_pp_group().is_last_rank):
+            self.wte = VocabParallelEmbedding(
+                self.vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                quant_config=quant_config,
+            )
+        else:
+            self.wte = PPMissingLayer()
+        self.start_layer, self.end_layer, self.h = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: ExaoneDecoderLayer(
+                config=config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=f"{prefix}.h",
+        )
+        if get_pp_group().is_last_rank:
+            self.ln_f = RMSNorm(config.hidden_size,
+                                eps=config.layer_norm_epsilon)
+        else:
+            self.ln_f = PPMissingLayer()
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.wte(input_ids)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.h[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+                residual,
+            )
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+
+        hidden_states, _ = self.ln_f(hidden_states, residual)
+        return hidden_states
+
+
+class ExaoneForCausalLM(nn.Module, SupportsLoRA):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "c_fc_0",
+            "c_fc_1",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj",
+        "out_proj",
+        "gate_up_proj",
+        "c_proj",
+        "wte",
+        "lm_head",
+    ]
+    embedding_modules = {
+        "wte": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+    embedding_padding_modules = ["lm_head"]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "c_fc_0": ("gate_up_proj", 0),
+        "c_fc_1": ("gate_up_proj", 1),
+    }
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        lora_config: Optional[LoRAConfig] = None,
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.lora_config = lora_config
+
+        self.transformer = ExaoneModel(
+            config,
+            cache_config,
+            quant_config,
+            lora_config=lora_config,
+            prefix="model",
+        )
+        if get_pp_group().is_last_rank:
+            self.unpadded_vocab_size = config.vocab_size
+            if lora_config:
+                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+            self.lm_head = ParallelLMHead(
+                self.unpadded_vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                padding_size=DEFAULT_VOCAB_PADDING_SIZE
+                # We need bigger padding if using lora for kernel
+                # compatibility
+                if not lora_config else lora_config.lora_vocab_padding_size,
+                quant_config=quant_config,
+            )
+            if config.tie_word_embeddings:
+                self.lm_head.weight = self.transformer.wte.weight
+
+            logit_scale = getattr(config, "logit_scale", 1.0)
+            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                    config.vocab_size,
+                                                    logit_scale)
+            self.sampler = Sampler()
+        else:
+            self.lm_head = PPMissingLayer()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        model_output = self.transformer(input_ids, positions, kv_caches,
+                                        attn_metadata, intermediate_tensors)
+        return model_output
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits
+
+    def sample(
+        self,
+        logits: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def make_empty_intermediate_tensors(
+            self, batch_size: int, dtype: torch.dtype,
+            device: torch.device) -> IntermediateTensors:
+        return IntermediateTensors({
+            "hidden_states":
+            torch.zeros(
+                (batch_size, self.config.hidden_size),
+                dtype=dtype,
+                device=device,
+            ),
+            "residual":
+            torch.zeros(
+                (batch_size, self.config.hidden_size),
+                dtype=dtype,
+                device=device,
+            ),
+        })
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        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", ".c_fc_0", 0),
+            (".gate_up_proj", ".c_fc_1", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in weights:
+            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
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+            if scale_name := get_compressed_tensors_cache_scale(name):
+                # Loading kv cache scales for compressed-tensors quantization
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = loaded_weight[0]
+                weight_loader(param, loaded_weight)
+                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
+
+                if is_pp_missing_parameter(name, self):
+                    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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+    # If this function is called, it should always initialize KV cache scale
+    # factors (or else raise an exception). Thus, handled exceptions should
+    # make sure to leave KV cache scale factors in a known good (dummy) state
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        for layer_idx, scaling_factor in kv_cache_scales_loader(
+                quantization_param_path,
+                tp_rank,
+                tp_size,
+                self.config.num_hidden_layers,
+                self.config.__class__.model_type,
+        ):
+            if not isinstance(self.transformer.h[layer_idx], nn.Identity):
+                layer_self_attn = self.transformer.h[layer_idx].attn
+
+            if is_hip():
+                # The scaling factor convention we are assuming is
+                # quantized_value * scaling_factor ~= true_value
+                # which is consistent with the practice of setting
+                # scaling_factor = tensor_amax / FPtype_max
+                scaling_factor *= 2
+            if hasattr(layer_self_attn, "kv_scale"):
+                layer_self_attn.attn._kv_scale = scaling_factor
+            else:
+                raise RuntimeError("Self attention has no KV cache scaling "
+                                   "factor attribute!")

vllm/model_executor/models/falcon.py

@@ -39,12 +39,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import RWConfig
 
 FalconConfig = Union[HF_FalconConfig, RWConfig]

vllm/model_executor/models/fuyu.py

@@ -31,6 +31,7 @@ from vllm.inputs import INPUT_REGISTRY, InputContext, LLMInputs
 from vllm.logger import init_logger
 from vllm.model_executor.layers.linear import ColumnParallelLinear
 from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.persimmon import PersimmonForCausalLM
 from vllm.model_executor.sampling_metadata import SamplingMetadata
@@ -39,7 +40,7 @@ from vllm.multimodal.base import MultiModalInputs
 from vllm.multimodal.image import cached_get_image_processor
 from vllm.multimodal.utils import cached_get_tokenizer
 from vllm.sequence import (VLLM_TOKEN_ID_ARRAY_TYPE, IntermediateTensors,
-                           SamplerOutput, SequenceData)
+                           SequenceData)
 
 from .interfaces import SupportsMultiModal
 from .utils import merge_multimodal_embeddings
@@ -249,6 +250,9 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal):
         image_patches = kwargs.pop("image_patches", None)
 
         if isinstance(image_patches, torch.Tensor):
+            # Remove the N dimension until multiple images are supported.
+            image_patches = image_patches.squeeze(1)
+
             expected_feature_size = self.image_feature_size
             if image_patches.size(-1) != expected_feature_size:
                 raise ValueError(

vllm/model_executor/models/gemma.py

@@ -34,12 +34,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsLoRA
 

vllm/model_executor/models/gemma2.py

@@ -33,12 +33,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsLoRA
 

vllm/model_executor/models/gpt2.py

@@ -34,12 +34,12 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 from .utils import is_pp_missing_parameter, make_layers
 

vllm/model_executor/models/gpt_bigcode.py

@@ -34,12 +34,12 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsLoRA
 

vllm/model_executor/models/gpt_j.py

@@ -33,12 +33,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 
 class GPTJAttention(nn.Module):

vllm/model_executor/models/gpt_neox.py

@@ -33,12 +33,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 
 class GPTNeoXAttention(nn.Module):

vllm/model_executor/models/granite.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 IBM Granite model compatible with HuggingFace weights."""
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.config import CacheConfig, LoRAConfig
+from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
+    get_compressed_tensors_cache_scale)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader, kv_cache_scales_loader, maybe_remap_kv_scale_name)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.configs.granite import GraniteConfig
+from vllm.utils import is_hip
+
+from .interfaces import SupportsLoRA
+from .utils import PPMissingLayer, is_pp_missing_parameter, make_layers
+
+
+class GraniteMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            input_size=hidden_size,
+            output_sizes=[intermediate_size] * 2,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj")
+        self.down_proj = RowParallelLinear(input_size=intermediate_size,
+                                           output_size=hidden_size,
+                                           bias=bias,
+                                           quant_config=quant_config,
+                                           prefix=f"{prefix}.down_proj")
+        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 GraniteAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: GraniteConfig,
+        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,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        cache_config: Optional[CacheConfig] = None,
+        prefix: str = "",
+    ) -> 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)
+        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+        self.head_dim = getattr(config, "head_dim",
+                                self.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 = config.attention_multiplier
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size=hidden_size,
+            head_size=self.head_dim,
+            total_num_heads=self.total_num_heads,
+            total_num_kv_heads=self.total_num_kv_heads,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.o_proj = RowParallelLinear(
+            input_size=self.total_num_heads * self.head_dim,
+            output_size=hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.o_proj",
+        )
+
+        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 = Attention(self.num_heads,
+                              self.head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_kv_heads,
+                              cache_config=cache_config,
+                              quant_config=quant_config)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class GraniteDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: GraniteConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.residual_multiplier = config.residual_multiplier
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+                config, "original_max_position_embeddings", None):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        # Support abacusai/Smaug-72B-v0.1 with attention_bias
+        # Support internlm/internlm-7b with bias
+        attention_bias = getattr(config, "attention_bias", False) or getattr(
+            config, "bias", False)
+        self.self_attn = GraniteAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=getattr(config, "num_key_value_heads",
+                                 config.num_attention_heads),
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            bias=attention_bias,
+            cache_config=cache_config,
+            prefix=f"{prefix}.self_attn",
+        )
+
+        self.mlp = GraniteMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            bias=getattr(config, "mlp_bias", False),
+            prefix=f"{prefix}.mlp",
+        )
+        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: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        hidden_states = residual + hidden_states * self.residual_multiplier
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states * self.residual_multiplier
+        return hidden_states
+
+
+class GraniteModel(nn.Module):
+
+    def __init__(
+        self,
+        config: GraniteConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        lora_config: Optional[LoRAConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        lora_vocab = (lora_config.lora_extra_vocab_size *
+                      (lora_config.max_loras or 1)) if lora_config else 0
+        self.vocab_size = config.vocab_size + lora_vocab
+        self.org_vocab_size = config.vocab_size
+        if get_pp_group().is_first_rank or (config.tie_word_embeddings
+                                            and get_pp_group().is_last_rank):
+            self.embed_tokens = VocabParallelEmbedding(
+                self.vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                quant_config=quant_config,
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: GraniteDecoderLayer(config=config,
+                                               cache_config=cache_config,
+                                               quant_config=quant_config,
+                                               prefix=prefix),
+            prefix=f"{prefix}.layers")
+        if get_pp_group().is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer()
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        hidden_states *= self.config.embedding_multiplier
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+            )
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+
+        hidden_states = self.norm(hidden_states)
+        return hidden_states
+
+
+class GraniteForCausalLM(nn.Module, SupportsLoRA):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj", "o_proj", "gate_up_proj", "down_proj", "embed_tokens",
+        "lm_head"
+    ]
+    embedding_modules = {
+        "embed_tokens": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+    embedding_padding_modules = ["lm_head"]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
+    def __init__(
+        self,
+        config: GraniteConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        lora_config: Optional[LoRAConfig] = None,
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.lora_config = lora_config
+
+        self.model = GraniteModel(config,
+                                  cache_config,
+                                  quant_config,
+                                  lora_config=lora_config,
+                                  prefix="model")
+        if get_pp_group().is_last_rank:
+            self.unpadded_vocab_size = config.vocab_size
+            if lora_config:
+                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+            self.lm_head = ParallelLMHead(
+                self.unpadded_vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                padding_size=DEFAULT_VOCAB_PADDING_SIZE
+                # We need bigger padding if using lora for kernel
+                # compatibility
+                if not lora_config else lora_config.lora_vocab_padding_size,
+                quant_config=quant_config,
+            )
+            if config.tie_word_embeddings:
+                self.lm_head.weight = self.model.embed_tokens.weight
+
+            logit_scale = getattr(config, "logit_scale", 1.0)
+            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                    config.vocab_size,
+                                                    logit_scale)
+            self.sampler = Sampler()
+        else:
+            self.lm_head = PPMissingLayer()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        model_output = self.model(input_ids, positions, kv_caches,
+                                  attn_metadata, intermediate_tensors)
+        return model_output
+
+    def compute_logits(
+            self, hidden_states: torch.Tensor,
+            sampling_metadata: SamplingMetadata) -> Optional[torch.Tensor]:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        logits /= self.config.logits_scaling
+        return logits
+
+    def sample(
+        self,
+        logits: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def make_empty_intermediate_tensors(
+            self, batch_size: int, dtype: torch.dtype,
+            device: torch.device) -> IntermediateTensors:
+        return IntermediateTensors({
+            "hidden_states":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+            "residual":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+        })
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        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 weights:
+            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
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+            if scale_name := get_compressed_tensors_cache_scale(name):
+                # Loading kv cache scales for compressed-tensors quantization
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = loaded_weight[0]
+                weight_loader(param, loaded_weight)
+                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
+
+                if is_pp_missing_parameter(name, self):
+                    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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+    # If this function is called, it should always initialize KV cache scale
+    # factors (or else raise an exception). Thus, handled exceptions should
+    # make sure to leave KV cache scale factors in a known good (dummy) state
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        for layer_idx, scaling_factor in kv_cache_scales_loader(
+                quantization_param_path, tp_rank, tp_size,
+                self.config.num_hidden_layers,
+                self.config.__class__.model_type):
+            if not isinstance(self.model.layers[layer_idx], nn.Identity):
+                layer_self_attn = self.model.layers[layer_idx].self_attn
+
+            if is_hip():
+                # The scaling factor convention we are assuming is
+                # quantized_value * scaling_factor ~= true_value
+                # which is consistent with the practice of setting
+                # scaling_factor = tensor_amax / FPtype_max
+                scaling_factor *= 2
+            if hasattr(layer_self_attn, "kv_scale"):
+                layer_self_attn.attn._kv_scale = scaling_factor
+            else:
+                raise RuntimeError("Self attention has no KV cache scaling "
+                                   "factor attribute!")

vllm/model_executor/models/intern_vit.py

@@ -11,13 +11,21 @@ import torch.nn as nn
 import torch.nn.functional as F
 from transformers import PretrainedConfig
 
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+try:
+    from xformers import ops as xops
+    USE_XFORMERS_OPS = True
+except ImportError:
+    USE_XFORMERS_OPS = False
+
 NORM2FN = {
     'rms_norm': RMSNorm,
     'layer_norm': nn.LayerNorm,
@@ -78,7 +86,73 @@ class InternVisionEmbeddings(nn.Module):
         return embeddings
 
 
-class InternAttention(nn.Module):
+class InternParallelAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f'embed_dim must be divisible by num_heads '
+                f'(got `embed_dim`: {self.embed_dim} and `num_heads`:'
+                f' {self.num_heads}).')
+
+        self.scale = self.head_dim**-0.5
+        self.qkv = QKVParallelLinear(
+            self.embed_dim,
+            self.head_dim,
+            self.num_heads,
+            bias=config.qkv_bias,
+            quant_config=quant_config,
+        )
+
+        self.qk_normalization = config.qk_normalization
+
+        if self.qk_normalization:
+            self.q_norm = RMSNorm(self.embed_dim, eps=config.layer_norm_eps)
+            self.k_norm = RMSNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+        self.proj = RowParallelLinear(
+            self.embed_dim,
+            self.embed_dim,
+            quant_config=quant_config,
+        )
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv, _ = self.qkv(x)
+        q, k, v = qkv.chunk(3, dim=-1)
+
+        q = q.view(B, N, self.num_heads_per_partition, self.head_dim)
+        k = k.view(B, N, self.num_heads_per_partition, self.head_dim)
+        v = v.view(B, N, self.num_heads_per_partition, self.head_dim)
+
+        if self.qk_normalization:
+            B_, N_, H_, D_ = q.shape
+            q = self.q_norm.forward_native(q.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+            k = self.k_norm.forward_native(k.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+
+        x = xops.memory_efficient_attention_forward(q, k, v, scale=self.scale)
+        x = x.view(B, N, -1)
+
+        x, _ = self.proj(x)
+        return x
+
+
+class InternSdpaAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
     def __init__(self, config: PretrainedConfig):
@@ -108,19 +182,25 @@ class InternAttention(nn.Module):
 
     def forward(self, x):
         B, N, C = x.shape
-        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads,
-                                  C // self.num_heads).permute(2, 0, 3, 1, 4)
-        q, k, v = qkv.unbind(0)
+        qkv = self.qkv(x)
+        q, k, v = qkv.chunk(3, dim=-1)
+
+        q = q.view(B, N, self.num_heads, self.head_dim)
+        k = k.view(B, N, self.num_heads, self.head_dim)
+        v = v.view(B, N, self.num_heads, self.head_dim)
 
         if self.qk_normalization:
-            B_, H_, N_, D_ = q.shape
-            q = self.q_norm.forward_native(q.transpose(1, 2).flatten(
-                -2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
-            k = self.k_norm.forward_native(k.transpose(1, 2).flatten(
-                -2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
+            B_, N_, H_, D_ = q.shape
+            q = self.q_norm.forward_native(q.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+            k = self.k_norm.forward_native(k.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
 
         x = F.scaled_dot_product_attention(q, k, v, scale=self.scale)
-        x = x.transpose(1, 2).reshape(B, N, C)
+        x = x.transpose(1, 2).view(B, N, -1)
 
         x = self.proj(x)
         return x
@@ -161,7 +241,14 @@ class InternVisionEncoderLayer(nn.Module):
         self.intermediate_size = config.intermediate_size
         self.norm_type = config.norm_type
 
-        self.attn = InternAttention(config)
+        # fallback to sdpa attention if tp unavailable
+        tp_size = get_tensor_model_parallel_world_size()
+        num_heads = config.num_attention_heads
+        if USE_XFORMERS_OPS and num_heads % tp_size == 0:
+            self.attn = InternParallelAttention(config,
+                                                quant_config=quant_config)
+        else:
+            self.attn = InternSdpaAttention(config)
         self.mlp = InternMLP(config, quant_config=quant_config)
         self.norm1 = NORM2FN[self.norm_type](self.embed_dim,
                                              eps=config.layer_norm_eps)

vllm/model_executor/models/internlm2.py

 # -*- coding: utf-8 -*-
+from functools import partial
 from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
@@ -7,7 +8,10 @@ from transformers import PretrainedConfig
 
 from vllm.attention import Attention, AttentionMetadata
 from vllm.config import CacheConfig
-from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.distributed import (get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size,
+                              split_tensor_along_last_dim,
+                              tensor_model_parallel_all_gather)
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
@@ -17,12 +21,12 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 
 class InternLM2MLP(nn.Module):
@@ -70,20 +74,21 @@ class InternLM2Attention(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
-        tp_size = get_tensor_model_parallel_world_size()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
         self.total_num_heads = num_heads
-        assert self.total_num_heads % tp_size == 0
-        self.num_heads = self.total_num_heads // tp_size
+        assert self.total_num_heads % self.tp_size == 0
+        self.num_heads = self.total_num_heads // self.tp_size
         self.total_num_kv_heads = num_kv_heads
-        if self.total_num_kv_heads >= tp_size:
+        if self.total_num_kv_heads >= self.tp_size:
             # Number of KV heads is greater than TP size, so we partition
             # the KV heads across multiple tensor parallel GPUs.
-            assert self.total_num_kv_heads % tp_size == 0
+            assert self.total_num_kv_heads % self.tp_size == 0
         else:
             # Number of KV heads is less than TP size, so we replicate
             # the KV heads across multiple tensor parallel GPUs.
-            assert tp_size % self.total_num_kv_heads == 0
-        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+            assert self.tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // self.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
@@ -122,11 +127,27 @@ class InternLM2Attention(nn.Module):
                               quant_config=quant_config)
 
     def split_qkv(self, qkv: torch.Tensor):
-        qkv = qkv.view(-1, self.num_kv_heads, self.key_value_groups + 2, 128)
-        q, k, v = torch.split(qkv, [self.key_value_groups, 1, 1], dim=2)
-        q = q.reshape(-1, self.q_size)
-        k = k.reshape(-1, self.kv_size)
-        v = v.reshape(-1, self.kv_size)
+        seq_len = qkv.shape[0]
+        if self.tp_size > 1:
+            qkv_map = [self.q_size, self.kv_size, self.kv_size] * self.tp_size
+            qkv = tensor_model_parallel_all_gather(qkv)
+            qkv = torch.split(qkv, qkv_map, dim=-1)
+            qkv = qkv[::3] + qkv[1::3] + qkv[2::3]
+            qkv = torch.cat(qkv, dim=-1)
+
+        qkv = qkv.view(seq_len, self.total_num_kv_heads,
+                       self.key_value_groups + 2, self.head_dim)
+        q, k, v = torch.split(qkv, [self.key_value_groups, 1, 1], dim=-2)
+        q = q.reshape(seq_len, self.q_size * self.tp_size)
+        k = k.reshape(seq_len, self.kv_size * self.tp_size)
+        v = v.reshape(seq_len, self.kv_size * self.tp_size)
+
+        if self.tp_size > 1:
+            splitter = partial(split_tensor_along_last_dim,
+                               num_partitions=self.tp_size)
+            q = splitter(q)[self.tp_rank]
+            k = splitter(k)[self.tp_rank]
+            v = splitter(v)[self.tp_rank]
         return q, k, v
 
     def forward(

vllm/model_executor/models/internvl.py

@@ -18,18 +18,19 @@ from vllm.attention import AttentionMetadata
 from vllm.config import CacheConfig, MultiModalConfig
 from vllm.inputs import INPUT_REGISTRY, InputContext, LLMInputs
 from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.intern_vit import InternVisionModel
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.base import MultiModalInputs
 from vllm.multimodal.utils import cached_get_tokenizer
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 
 from .clip import (dummy_image_for_clip, dummy_seq_data_for_clip,
                    get_clip_num_patches)
 from .interfaces import SupportsMultiModal
-from .utils import (filter_weights, init_vllm_registered_model,
+from .utils import (filter_weights, flatten_bn, init_vllm_registered_model,
                     merge_multimodal_embeddings)
 
 IMG_START = '<img>'
@@ -42,19 +43,17 @@ IMAGENET_STD = (0.229, 0.224, 0.225)
 
 class InternVLImagePixelInputs(TypedDict):
     type: Literal["pixel_values"]
-    data: Union[torch.Tensor, List[torch.Tensor]]
+    data: torch.Tensor
     """
-    Shape: `(batch_size, 1 + num_patches, num_channels, height, width)`
-
-    Note that `num_patches` may be different for each batch, in which case
-    the data is passed as a list instead of a batched tensor.
+    Shape:
+    `(batch_size * num_images * (1 + num_patches), num_channels, height, width)`
     """
 
 
 class InternVLImageEmbeddingInputs(TypedDict):
     type: Literal["image_embeds"]
-    data: Union[torch.Tensor, List[torch.Tensor]]
-    """Shape: `(batch_size, image_feature_size, hidden_size)`
+    data: torch.Tensor
+    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
 
     `hidden_size` must match the hidden size of language model backbone.
     """
@@ -244,6 +243,8 @@ def input_mapper_for_internvl(ctx: InputContext, data: object):
                                      min_num,
                                      max_num,
                                      use_thumbnail=use_thumbnail)
+        # Add an N dimension for number of images per prompt (currently 1).
+        data = data.unsqueeze(0)
     model_config = ctx.model_config
     tokenizer = cached_get_tokenizer(model_config.tokenizer,
                                      trust_remote_code=True)
@@ -355,7 +356,7 @@ class InternVLChatModel(nn.Module, SupportsMultiModal):
             x = x.permute(0, 2, 1, 3).contiguous()
         return x
 
-    def extract_feature(self, pixel_values):
+    def extract_feature(self, pixel_values: torch.Tensor) -> torch.Tensor:
         vit_embeds = self.vision_model(pixel_values=pixel_values)
         vit_embeds = vit_embeds[:, 1:, :]
 
@@ -368,17 +369,7 @@ class InternVLChatModel(nn.Module, SupportsMultiModal):
         vit_embeds = self.mlp1(vit_embeds)
         return vit_embeds
 
-    def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
-        if list(data.shape[1:]) != [2]:
-            raise ValueError(
-                f"The expected image sizes shape is batch dimension plus "
-                f"{[2]}. You supplied {data.shape}.")
-
-        return data
-
-    def _validate_pixel_values(
-        self, data: Union[torch.Tensor, List[torch.Tensor]]
-    ) -> Union[torch.Tensor, List[torch.Tensor]]:
+    def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
 
         h = w = self.config.vision_config.image_size
         expected_dims = (3, h, w)
@@ -387,10 +378,11 @@ class InternVLChatModel(nn.Module, SupportsMultiModal):
             actual_dims = tuple(d.shape)
 
             if actual_dims != expected_dims:
-                expected_expr = ("num_patches", *map(str, expected_dims))
+                expected_expr = str(expected_dims)
                 raise ValueError(
-                    "The expected shape of pixel values in each batch element "
-                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
+                    "The expected shape of pixel values per image per batch "
+                    f" per patch is {expected_expr}. "
+                    f"You supplied {tuple(d.shape)}.")
 
         for d in data:
             _validate_shape(d)
@@ -410,9 +402,10 @@ class InternVLChatModel(nn.Module, SupportsMultiModal):
             if not isinstance(image_embeds, torch.Tensor):
                 raise ValueError("Incorrect type of image embeddings. "
                                  f"Got type: {type(image_embeds)}")
+
             return InternVLImageEmbeddingInputs(
                 type="image_embeds",
-                data=image_embeds,
+                data=flatten_bn(image_embeds),
             )
 
         self.img_context_token_id = image_token_id[0]
@@ -424,7 +417,8 @@ class InternVLChatModel(nn.Module, SupportsMultiModal):
 
             return InternVLImagePixelInputs(
                 type="pixel_values",
-                data=self._validate_pixel_values(pixel_values),
+                data=self._validate_pixel_values(
+                    flatten_bn(pixel_values, concat=True).flatten(0, 1)),
             )
 
         raise AssertionError("This line should be unreachable.")

vllm/model_executor/models/jais.py

@@ -35,12 +35,12 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import JAISConfig
 
 from .utils import is_pp_missing_parameter, make_layers

vllm/model_executor/models/jamba.py

@@ -4,9 +4,6 @@ from dataclasses import dataclass
 from typing import Dict, Iterable, List, Optional, Tuple
 
 import torch
-from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
-from mamba_ssm.ops.selective_scan_interface import selective_scan_fn
-from mamba_ssm.ops.triton.selective_state_update import selective_state_update
 from torch import nn
 from torch.nn.parameter import Parameter
 from transformers import JambaConfig
@@ -24,16 +21,20 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                ReplicatedLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
+    causal_conv1d_fn, causal_conv1d_update)
+from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
+    selective_scan_fn, selective_state_update)
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.interfaces import HasInnerState
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 from vllm.worker.model_runner import (_BATCH_SIZES_TO_CAPTURE,
                                       _get_graph_batch_size)
 
@@ -161,7 +162,7 @@ class JambaMambaMixer(nn.Module):
                     (self.conv_kernel_size - hidden_states.shape[-1], 0))
                 cache_params.conv_state.copy_(conv_states)
 
-            hidden_states = causal_conv1d_fn(
+            hidden_states, _ = causal_conv1d_fn(
                 hidden_states,
                 conv_weights,
                 self.conv1d.bias,
@@ -920,7 +921,7 @@ class JambaForCausalLM(nn.Module, HasInnerState):
                     weight_loader = param.weight_loader
                     weight_loader(param,
                                   loaded_weight,
-                                  weight_name,
+                                  name,
                                   shard_id=shard_id,
                                   expert_id=expert_id)
                     break

vllm/model_executor/models/llama.py

@@ -44,13 +44,13 @@ from vllm.model_executor.layers.quantization.base_config import (
 from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
     get_compressed_tensors_cache_scale)
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import (
     default_weight_loader, kv_cache_scales_loader, maybe_remap_kv_scale_name)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import IntermediateTensors, SamplerOutput
+from vllm.sequence import IntermediateTensors
 from vllm.utils import is_hip
 
 from .interfaces import SupportsLoRA