Gemma Support (#256)

python/sglang/lang/chat_template.py

@@ -17,15 +17,23 @@ class ChatTemplate:
     image_token: str = "<image>"
     style: ChatTemplateStyle = ChatTemplateStyle.PLAIN
 
-    def get_prefix_and_suffix(self, role: str, hist_messages: List[Dict]) -> Tuple[str, str]:
+    def get_prefix_and_suffix(
+        self, role: str, hist_messages: List[Dict]
+    ) -> Tuple[str, str]:
         prefix, suffix = self.role_prefix_and_suffix.get(role, ("", ""))
-        
+
         if self.style == ChatTemplateStyle.LLAMA2:
             if role == "system" and not hist_messages:
                 user_prefix, _ = self.role_prefix_and_suffix.get("user", ("", ""))
-                system_prefix, system_suffix = self.role_prefix_and_suffix.get("system", ("", ""))
+                system_prefix, system_suffix = self.role_prefix_and_suffix.get(
+                    "system", ("", "")
+                )
                 return (user_prefix + system_prefix, system_suffix)
-            elif role == "user" and len(hist_messages) == 1 and hist_messages[0]["content"] is not None:
+            elif (
+                role == "user"
+                and len(hist_messages) == 1
+                and hist_messages[0]["content"] is not None
+            ):
                 return ("", suffix)
 
         return prefix, suffix
@@ -171,6 +179,19 @@ register_chat_template(
     )
 )
 
+register_chat_template(
+    ChatTemplate(
+        name="gemma-it",
+        default_system_prompt=None,
+        role_prefix_and_suffix={
+            "system": ("", ""),
+            "user": ("<start_of_turn>user\n", "<end_of_turn>\n"),
+            "assistant": ("<start_of_turn>model\n", "<end_of_turn>\n"),
+        },
+        style=ChatTemplateStyle.PLAIN,
+    )
+)
+
 
 @register_chat_template_matching_function
 def match_vicuna(model_path: str):
@@ -211,6 +232,13 @@ def match_chat_yi(model_path: str):
         return get_chat_template("yi")
 
 
+@register_chat_template_matching_function
+def match_gemma_it(model_path: str):
+    model_path = model_path.lower()
+    if "gemma" in model_path and "it" in model_path:
+        return get_chat_template("gemma-it")
+
+
 if __name__ == "__main__":
     messages = [
         {"role": "system", "content": None},  # None means default

python/sglang/srt/layers/context_flashattention_nopad.py

@@ -129,7 +129,7 @@ def context_attention_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len):
 
     Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
     assert Lq == Lk and Lk == Lv
-    assert Lk in {16, 32, 64, 128}
+    assert Lk in {16, 32, 64, 128, 256}
 
     sm_scale = 1.0 / (Lq**0.5)
     batch, head = b_seq_len.shape[0], q.shape[1]

python/sglang/srt/layers/extend_attention.py

@@ -181,19 +181,20 @@ def extend_attention_fwd(
 
     k_buffer, v_buffer: (prefix + extend) tensors in mem_manager
     """
-    if CUDA_CAPABILITY[0] >= 8:
-        BLOCK_M, BLOCK_N = 128, 128
-    else:
-        BLOCK_M, BLOCK_N = 64, 64
-
     Lq, Lk, Lv, Lo = (
         q_extend.shape[-1],
         k_extend.shape[-1],
         v_extend.shape[-1],
         o_extend.shape[-1],
     )
+
     assert Lq == Lk and Lk == Lv and Lv == Lo
-    assert Lq in {16, 32, 64, 128}
+    assert Lq in {16, 32, 64, 128, 256}
+
+    if CUDA_CAPABILITY[0] >= 8:
+        BLOCK_M, BLOCK_N = (128, 128) if Lq <= 128 else (64, 64)
+    else:
+        BLOCK_M, BLOCK_N = (64, 64) if Lq <= 128 else (32, 32)
 
     sm_scale = 1.0 / (Lq**0.5)
     batch_size, head_num = b_seq_len.shape[0], q_extend.shape[1]

python/sglang/srt/layers/radix_attention.py

@@ -1,15 +1,9 @@
-from typing import List
-
 import torch
 from sglang.srt.layers.context_flashattention_nopad import context_attention_fwd
 from sglang.srt.layers.extend_attention import extend_attention_fwd
 from sglang.srt.layers.token_attention import token_attention_fwd
 from sglang.srt.managers.router.model_runner import ForwardMode, InputMetadata
 from torch import nn
-from vllm.model_executor.parallel_utils.parallel_state import (
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-)
 
 
 class RadixAttention(nn.Module):
@@ -21,9 +15,9 @@ class RadixAttention(nn.Module):
         self.head_dim = head_dim
         self.layer_id = layer_id
 
-        from sglang.srt.managers.router.model_runner import global_model_mode
+        from sglang.srt.managers.router.model_runner import global_server_args
 
-        self.use_flashinfer = "flashinfer" in global_model_mode
+        self.use_flashinfer = "flashinfer" in global_server_args.model_mode
 
         if self.use_flashinfer:
             self.prefill_forward = self.prefill_forward_flashinfer

python/sglang/srt/layers/token_attention.py

@@ -5,6 +5,14 @@ import torch
 import triton
 import triton.language as tl
 from sglang.srt.utils import wrap_kernel_launcher
+from sglang.srt.managers.router.model_runner import global_server_args
+
+if global_server_args.attention_reduce_in_fp32:
+    REDUCE_TRITON_TYPE = tl.float32
+    REDUCE_TORCH_TYPE = torch.float32
+else:
+    REDUCE_TRITON_TYPE = tl.float16
+    REDUCE_TORCH_TYPE = torch.float16
 
 
 @triton.jit
@@ -49,7 +57,7 @@ def _fwd_kernel_stage1(
     block_mask = tl.where(block_stard_index < cur_batch_seq_len, 1, 0)
 
     for start_mark in range(0, block_mask, 1):
-        q = tl.load(Q + off_q + start_mark)
+        q = tl.load(Q + off_q + start_mark).to(REDUCE_TRITON_TYPE)
         offs_n_new = cur_batch_start_index + offs_n
         k_loc = tl.load(
             Req_to_tokens + stride_req_to_tokens_b * cur_batch_req_idx + offs_n_new,
@@ -65,7 +73,7 @@ def _fwd_kernel_stage1(
             K_Buffer + offs_buf_k,
             mask=offs_n_new[:, None] < cur_batch_end_index,
             other=0.0,
-        )
+        ).to(REDUCE_TRITON_TYPE)
         att_value = tl.sum(q[None, :] * k, 1)
         att_value *= sm_scale
         off_o = cur_head * att_stride_h + (cur_batch_in_all_start_index + offs_n)
@@ -161,7 +169,7 @@ def _token_att_m_fwd(
     # shape constraints
     Lq, Lk = q.shape[-1], k_buffer.shape[-1]
     assert Lq == Lk
-    assert Lk in {16, 32, 64, 128}
+    assert Lk in {16, 32, 64, 128, 256}
     sm_scale = 1.0 / (Lk**0.5)
 
     batch, head_num = B_req_idx.shape[0], q.shape[1]
@@ -299,7 +307,7 @@ def token_attention_fwd(
 ):
     if att_m is None:
         att_m = torch.empty(
-            (q.shape[-2], total_num_tokens), dtype=q.dtype, device="cuda"
+            (q.shape[-2], total_num_tokens), dtype=REDUCE_TORCH_TYPE, device="cuda"
         )
 
     _token_att_m_fwd(

python/sglang/srt/managers/router/model_rpc.py

@@ -57,17 +57,19 @@ class ModelRpcServer(rpyc.Service):
 
         # Init model and tokenizer
         self.model_config = ModelConfig(
-            server_args.model_path, server_args.trust_remote_code, context_length=server_args.context_length
+            server_args.model_path,
+            server_args.trust_remote_code,
+            context_length=server_args.context_length,
         )
         self.model_runner = ModelRunner(
-            self.model_config,
-            server_args.mem_fraction_static,
-            tp_rank,
-            server_args.tp_size,
-            port_args.nccl_port,
-            server_args.load_format,
-            server_args.trust_remote_code,
-            server_args.model_mode,
+            model_config=self.model_config,
+            mem_fraction_static=server_args.mem_fraction_static,
+            tp_rank=tp_rank,
+            tp_size=server_args.tp_size,
+            nccl_port=port_args.nccl_port,
+            server_args=server_args,
+            load_format=server_args.load_format,
+            trust_remote_code=server_args.trust_remote_code,
         )
         if is_multimodal_model(server_args.model_path):
             self.processor = get_processor(
@@ -435,7 +437,7 @@ class ModelRpcServer(rpyc.Service):
                 # If logprob_start_len > 0, then first logprob_start_len prompt tokens
                 # will be ignored.
                 prompt_token_len = len(req.logprob)
-                token_ids = req.input_ids[-prompt_token_len :] + [next_token_ids[i]]
+                token_ids = req.input_ids[-prompt_token_len:] + [next_token_ids[i]]
                 token_logprobs = req.logprob + [last_logprobs[i]]
                 req.token_logprob = list(zip(token_ids, token_logprobs))
                 if req.logprob_start_len == 0:
@@ -553,8 +555,7 @@ class ModelRpcServer(rpyc.Service):
                     "completion_tokens": len(req.input_ids)
                     + len(req.output_ids)
                     - req.prompt_tokens,
-                    "completion_tokens_wo_jump_forward":
-                    req.completion_tokens_wo_jump_forward
+                    "completion_tokens_wo_jump_forward": req.completion_tokens_wo_jump_forward,
                 }
                 if req.return_logprob:
                     meta_info["prompt_logprob"] = req.logprob

python/sglang/srt/managers/router/model_runner.py

@@ -3,7 +3,6 @@ import logging
 from dataclasses import dataclass
 from functools import lru_cache
 from pathlib import Path
-from typing import List
 
 import numpy as np
 import torch
@@ -23,8 +22,8 @@ QUANTIONCONFIG_MAPPING = {"awq": AWQConfig, "gptq": GPTQConfig}
 logger = logging.getLogger("model_runner")
 
 
-# for model_mode
-global_model_mode: List[str] = []
+# for server args in model endpoints
+global_server_args = None
 
 
 @lru_cache()
@@ -81,7 +80,6 @@ class InputMetadata:
     return_logprob: bool = False
 
     # for flashinfer
-    use_flashinfer: bool = False
     qo_indptr: torch.Tensor = None
     kv_indptr: torch.Tensor = None
     kv_indices: torch.Tensor = None
@@ -224,8 +222,7 @@ class InputMetadata:
         if forward_mode == ForwardMode.EXTEND:
             ret.init_extend_args()
 
-        ret.use_flashinfer = "flashinfer" in model_runner.model_mode
-        if ret.use_flashinfer:
+        if "flashinfer" in global_server_args.model_mode:
             ret.init_flashinfer_args(tp_size)
 
         return ret
@@ -239,9 +236,9 @@ class ModelRunner:
         tp_rank,
         tp_size,
         nccl_port,
+        server_args,
         load_format="auto",
         trust_remote_code=True,
-        model_mode: List[str] = (),
     ):
         self.model_config = model_config
         self.mem_fraction_static = mem_fraction_static
@@ -250,10 +247,9 @@ class ModelRunner:
         self.nccl_port = nccl_port
         self.load_format = load_format
         self.trust_remote_code = trust_remote_code
-        self.model_mode = model_mode
 
-        global global_model_mode
-        global_model_mode = model_mode
+        global global_server_args
+        global_server_args = server_args
 
         # Init torch distributed
         torch.cuda.set_device(self.tp_rank)
@@ -319,9 +315,7 @@ class ModelRunner:
         available_gpu_memory = get_available_gpu_memory(
             self.tp_rank, distributed=self.tp_size > 1
         ) * (1 << 30)
-        head_dim = (
-            self.model_config.hidden_size // self.model_config.num_attention_heads
-        )
+        head_dim = self.model_config.head_dim
         head_num = self.model_config.num_key_value_heads // self.tp_size
         cell_size = head_num * head_dim * self.model_config.num_hidden_layers * 2 * 2
         rest_memory = available_gpu_memory - total_gpu_memory * (
@@ -346,8 +340,7 @@ class ModelRunner:
             self.max_total_num_token,
             dtype=torch.float16,
             head_num=self.model_config.num_key_value_heads // self.tp_size,
-            head_dim=self.model_config.hidden_size
-            // self.model_config.num_attention_heads,
+            head_dim=self.model_config.head_dim,
             layer_num=self.model_config.num_hidden_layers,
         )
 

python/sglang/srt/model_config.py

@@ -1,7 +1,5 @@
-import os
-from typing import Optional, Union
+from typing import Optional
 
-import torch
 from sglang.srt.hf_transformers_utils import get_config, get_context_length
 
 
@@ -17,14 +15,18 @@ class ModelConfig:
         self.trust_remote_code = trust_remote_code
         self.revision = revision
         self.hf_config = get_config(self.path, trust_remote_code, revision)
-        
+
         if context_length is not None:
             self.context_len = context_length
         else:
             self.context_len = get_context_length(self.hf_config)
 
         # Unify the config keys for hf_config
-        self.head_dim = self.hf_config.hidden_size // self.hf_config.num_attention_heads
+        self.head_dim = getattr(
+            self.hf_config,
+            "head_dim",
+            self.hf_config.hidden_size // self.hf_config.num_attention_heads,
+        )
         self.num_attention_heads = self.hf_config.num_attention_heads
         self.num_key_value_heads = getattr(self.hf_config, "num_key_value_heads", None)
         if self.num_key_value_heads is None:

python/sglang/srt/models/gemma.py

@@ -0,0 +1,340 @@
+# Adapted from:
+# https://github.com/vllm-project/vllm/blob/d65fac2738f0287a41955b45df76a2d5a919bff6/vllm/model_executor/models/gemma.py
+"""Inference-only Gemma model compatible with HuggingFace weights."""
+from typing import Optional, Tuple
+
+import torch
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from torch import nn
+from transformers import GemmaConfig
+from vllm.config import LoRAConfig
+from vllm.model_executor.input_metadata import InputMetadata
+from vllm.model_executor.layers.activation import GeluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    LinearMethodBase,
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from vllm.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_world_size,
+)
+from vllm.model_executor.weight_utils import (
+    default_weight_loader,
+    hf_model_weights_iterator,
+)
+
+
+class GemmaMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size, hidden_size, bias=False, linear_method=linear_method
+        )
+        self.act_fn = GeluAndMul()
+
+    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 GemmaAttention(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        layer_id: int = 0,
+        max_position_embeddings: int = 8192,
+        rope_theta: float = 10000,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = head_dim
+        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.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            linear_method=linear_method,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            linear_method=linear_method,
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=self.rope_theta,
+            is_neox_style=True,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class GemmaDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: GemmaConfig,
+        layer_id: int = 0,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = GemmaAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            head_dim=config.head_dim,
+            layer_id=layer_id,
+            max_position_embeddings=config.max_position_embeddings,
+            rope_theta=config.rope_theta,
+            linear_method=linear_method,
+        )
+        self.mlp = GemmaMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            linear_method=linear_method,
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class GemmaModel(nn.Module):
+    def __init__(
+        self,
+        config: GemmaConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                GemmaDecoderLayer(config, i, linear_method)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        skip_embed: bool = False,
+    ) -> torch.Tensor:
+        if not skip_embed:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_ids
+
+        # Normalize the embedding by sqrt(hidden_size)
+        hidden_states *= self.config.hidden_size**0.5
+
+        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 GemmaForCausalLM(nn.Module):
+    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",
+    ]
+    # Gemma does not apply LoRA to the embedding layer.
+    embedding_modules = {}
+    embedding_padding_modules = []
+
+    def __init__(
+        self,
+        config: GemmaConfig,
+        linear_method: Optional[LinearMethodBase] = None,
+        lora_config: Optional[LoRAConfig] = None,
+    ) -> None:
+        del lora_config  # Unused.
+        super().__init__()
+        self.config = config
+        self.linear_method = linear_method
+        self.model = GemmaModel(config, linear_method)
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        skip_embed: bool = False,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata, skip_embed)
+        return self.logits_processor(
+            input_ids, hidden_states, self.model.embed_tokens.weight, input_metadata
+        )
+
+    def load_weights(
+        self,
+        model_name_or_path: str,
+        cache_dir: Optional[str] = None,
+        load_format: str = "auto",
+        revision: Optional[str] = None,
+    ):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+        loaded_params = set()
+        for name, loaded_weight in hf_model_weights_iterator(
+            model_name_or_path, cache_dir, load_format, revision
+        ):
+            for param_name, shard_name, shard_id in stacked_params_mapping:
+                if shard_name not in name:
+                    continue
+                name = name.replace(shard_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                # GemmaRMSNorm is different from Llama's in that it multiplies
+                # (1 + weight) to the output, instead of just weight.
+                if "norm.weight" in name:
+                    loaded_weight += 1.0
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        unloaded_params = params_dict.keys() - loaded_params
+        if unloaded_params:
+            raise RuntimeError(
+                "Some weights are not initialized from checkpoints: "
+                f"{unloaded_params}"
+            )
+
+
+EntryClass = GemmaForCausalLM

python/sglang/srt/server_args.py

@@ -28,6 +28,7 @@ class ServerArgs:
     log_level: str = "info"
     disable_regex_jump_forward: bool = False
     disable_disk_cache: bool = False
+    attention_reduce_in_fp32: bool = False
 
     def __post_init__(self):
         if self.tokenizer_path is None:
@@ -189,6 +190,11 @@ class ServerArgs:
             action="store_true",
             help="Disable disk cache to avoid possible crashes related to file system or high concurrency.",
         )
+        parser.add_argument(
+            "--attention-reduce-in-fp32",
+            action="store_true",
+            help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16.",
+        )
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):