sglang 0.1.22__py3-none-any.whl → 0.1.25__py3-none-any.whl

sglang/srt/models/deepseek.py

@@ -0,0 +1,430 @@
+# Adapted from:
+# https://github.com/vllm-project/vllm/blob/14f91fe67c2342f2fe859dc6a5c40810df0e1c61/vllm/model_executor/models/deepseek.py
+"""Inference-only Deepseek model."""
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+from vllm.config import CacheConfig
+from vllm.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_reduce,
+)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import fused_moe
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
+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.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.controller.infer_batch import InputMetadata
+
+
+class DeepseekMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        reduce_results: bool = True,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            reduce_results=reduce_results,
+        )
+        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 DeepseekMoE(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.rank = get_tensor_model_parallel_rank()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.n_routed_experts = config.n_routed_experts
+        self.top_k = config.num_experts_per_tok
+        if self.tp_size > self.n_routed_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {self.n_routed_experts}."
+            )
+
+        self.experts = nn.ModuleList(
+            [
+                DeepseekMLP(
+                    hidden_size=config.hidden_size,
+                    intermediate_size=config.moe_intermediate_size,
+                    hidden_act=config.hidden_act,
+                    quant_config=quant_config,
+                    reduce_results=False,
+                )
+                for idx in range(self.n_routed_experts)
+            ]
+        )
+        self.pack_params()
+
+        self.gate = ReplicatedLinear(
+            config.hidden_size, self.n_routed_experts, bias=False, quant_config=None
+        )
+
+        if config.n_shared_experts is not None:
+            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
+            self.shared_experts = DeepseekMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+            )
+
+    def pack_params(self):
+        w1 = []
+        w2 = []
+        for expert in self.experts:
+            w1.append(expert.gate_up_proj.weight)
+            w2.append(expert.down_proj.weight)
+        self.w1 = torch._utils._flatten_dense_tensors(w1)
+        w1s = torch._utils._unflatten_dense_tensors(self.w1, w1)
+        for data, param in zip(w1s, w1):
+            param.data = data
+        self.w1 = self.w1.view(len(w1), *w1s[0].shape)
+
+        self.w2 = torch._utils._flatten_dense_tensors(w2)
+        w2s = torch._utils._unflatten_dense_tensors(self.w2, w2)
+        for data, param in zip(w2s, w2):
+            param.data = data
+
+        self.w2 = self.w2.view(len(w2), *w2s[0].shape)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        if self.config.n_shared_experts is not None:
+            shared_output = self.shared_experts(hidden_states)
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.gate(hidden_states)
+        final_hidden_states = fused_moe(
+            hidden_states,
+            self.w1,
+            self.w2,
+            router_logits,
+            self.top_k,
+            renormalize=self.config.norm_topk_prob,
+            inplace=True,
+        )
+
+        if self.config.n_shared_experts is not None:
+            final_hidden_states = final_hidden_states + shared_output
+        final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
+
+        return final_hidden_states.view(num_tokens, hidden_dim)
+
+
+class DeepseekAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = 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,
+            quant_config=quant_config,
+        )
+
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class DeepseekDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = 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 = DeepseekAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+        )
+        if (
+            config.n_routed_experts is not None
+            and layer_id >= config.first_k_dense_replace
+            and layer_id % config.moe_layer_freq == 0
+        ):
+            self.mlp = DeepseekMoE(config=config, quant_config=quant_config)
+        else:
+            self.mlp = DeepseekMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+            )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class DeepseekModel(nn.Module):
+
+    fall_back_to_pt_during_load = False
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = 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(
+            [
+                DeepseekDecoderLayer(
+                    config, layer_id, cache_config, quant_config=quant_config
+                )
+                for layer_id in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> 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, input_metadata, residual
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class DeepseekForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = DeepseekModel(config, cache_config, quant_config)
+        self.lm_head = ParallelLMHead(
+            config.vocab_size, config.hidden_size, quant_config=quant_config
+        )
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    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
+            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
+                # Skip experts that are not assigned to this worker.
+                if (
+                    "mlp.experts." in name or "mlp.shared_experts." in name
+                ) 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 (
+                    "mlp.experts." in name or "mlp.shared_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)
+
+
+EntryClass = DeepseekForCausalLM

sglang/srt/models/gpt_bigcode.py

@@ -0,0 +1,282 @@
+# Adapted from:
+# https://github.com/vllm-project/vllm/blob/07eb6f19f3b0ee9f7adf6eb689607028aa40bfd5/vllm/model_executor/models/gpt_bigcode.py
+"""Inference-only GPTBigCode model compatible with HuggingFace weights."""
+from typing import Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import GPTBigCodeConfig
+from vllm.config import CacheConfig, LoRAConfig
+from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.controller.infer_batch import InputMetadata
+
+
+class GPTBigCodeAttention(nn.Module):
+
+    def __init__(
+        self,
+        layer_id: int,
+        config: GPTBigCodeConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        total_num_heads = config.num_attention_heads
+        self.tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
+        assert total_num_heads % self.tensor_model_parallel_world_size == 0
+        self.num_heads = total_num_heads // self.tensor_model_parallel_world_size
+        self.head_dim = self.hidden_size // total_num_heads
+        self.scale = self.head_dim**-0.5
+
+        self.multi_query = config.multi_query
+        if self.multi_query:
+            total_num_kv_heads = 1
+            self.num_kv_heads = 1
+        else:
+            total_num_kv_heads = total_num_heads
+            self.num_kv_heads = self.num_heads
+        self.kv_dim = self.head_dim * self.num_kv_heads
+        self.c_attn = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            total_num_heads,
+            total_num_kv_heads,
+            bias=True,
+            quant_config=quant_config,
+        )
+
+        self.c_proj = RowParallelLinear(
+            self.hidden_size,
+            self.hidden_size,
+            bias=True,
+            quant_config=quant_config,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            scaling=self.scale,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.c_attn(hidden_states)
+        q, k, v = qkv.split(
+            [
+                self.hidden_size // self.tensor_model_parallel_world_size,
+                self.kv_dim,
+                self.kv_dim,
+            ],
+            dim=-1,
+        )
+        attn_output = self.attn(q, k, v, input_metadata)
+        attn_output, _ = self.c_proj(attn_output)
+        return attn_output
+
+
+class GPTBigMLP(nn.Module):
+
+    def __init__(
+        self,
+        intermediate_size: int,
+        config: GPTBigCodeConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.c_fc = ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            bias=True,
+            quant_config=quant_config,
+        )
+        self.c_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=True,
+            quant_config=quant_config,
+        )
+        self.act = get_act_fn(
+            config.activation_function, quant_config, intermediate_size
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.c_proj(hidden_states)
+        return hidden_states
+
+
+class GPTBigCodeBlock(nn.Module):
+
+    def __init__(
+        self,
+        layer_id: int,
+        config: GPTBigCodeConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
+
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = GPTBigCodeAttention(layer_id, config, cache_config, quant_config)
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.mlp = GPTBigMLP(inner_dim, config, quant_config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_output = self.attn(
+            hidden_states=hidden_states, input_metadata=input_metadata
+        )
+        # residual connection
+        hidden_states = attn_output + residual
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+        return hidden_states
+
+
+class GPTBigCodeModel(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTBigCodeConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        lora_config: Optional[LoRAConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        assert not config.add_cross_attention
+
+        self.embed_dim = config.hidden_size
+        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 = VocabParallelEmbedding(
+            self.vocab_size, self.embed_dim, org_num_embeddings=config.vocab_size
+        )
+        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
+        self.h = nn.ModuleList(
+            [
+                GPTBigCodeBlock(i, config, cache_config, quant_config)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+
+        for i in range(len(self.h)):
+            layer = self.h[i]
+            hidden_states = layer(hidden_states, input_metadata)
+
+        hidden_states = self.ln_f(hidden_states)
+        return hidden_states
+
+
+class GPTBigCodeForCausalLM(nn.Module):
+    packed_modules_mapping = {"c_attn": ["c_attn"]}
+
+    supported_lora_modules = ["c_fc", "c_proj", "wte", "c_attn"]
+
+    embedding_modules = {
+        "wte": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+
+    embedding_padding_modules = []
+
+    def __init__(
+        self,
+        config: GPTBigCodeConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        lora_config: Optional[LoRAConfig] = None,
+    ):
+        super().__init__()
+
+        self.config = config
+        self.lora_config = lora_config
+
+        self.quant_config = quant_config
+        self.transformer = GPTBigCodeModel(
+            config, cache_config, quant_config, lora_config
+        )
+        self.lm_head = self.transformer.wte
+        self.unpadded_vocab_size = config.vocab_size
+        if lora_config:
+            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.transformer(input_ids, positions, input_metadata)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        for name, loaded_weight in weights:
+            if "lm_head.weight" in name:
+                continue
+            if ".attn.bias" in name:
+                # Skip attention mask.
+                # NOTE: "c_attn.bias" should not be skipped.
+                continue
+            param = params_dict[name]
+            weight_loader = getattr(param, "weight_loader", default_weight_loader)
+            # TODO (@robertgshaw2-neuralmagic): move to fp8 linear method
+            if "c_attn.input_scale" in name or "c_attn.weight_scale" in name:
+                weight_loader(param, loaded_weight, "q")
+                weight_loader(param, loaded_weight, "k")
+                weight_loader(param, loaded_weight, "v")
+            else:
+                weight_loader(param, loaded_weight)
+
+
+EntryClass = GPTBigCodeForCausalLM