Expose dtype argument (#569)

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

@@ -6,7 +6,7 @@ import logging
 import pkgutil
 from dataclasses import dataclass
 from functools import lru_cache
-from typing import List, Optional, Type, Any
+from typing import List, Optional, Type
 
 import numpy as np
 import torch
@@ -119,7 +119,7 @@ class InputMetadata:
                 head_dim,
                 1,
                 pos_encoding_mode="NONE",
-                data_type="float16",
+                data_type=self.token_to_kv_pool.kv_data[0].dtype
             )
 
     def init_extend_args(self):
@@ -287,10 +287,11 @@ class ModelRunner:
             tokenizer=None,
             tokenizer_mode=None,
             trust_remote_code=self.server_args.trust_remote_code,
-            dtype=torch.float16,
+            dtype=self.server_args.dtype,
             seed=42,
             skip_tokenizer_init=True,
         )
+        self.dtype = vllm_model_config.dtype
         if self.model_config.model_overide_args is not None:
             vllm_model_config.hf_config.update(self.model_config.model_overide_args)
 
@@ -307,6 +308,7 @@ class ModelRunner:
         logger.info(
             f"[gpu_id={self.gpu_id}] Load weight end. "
             f"type={type(self.model).__name__}, "
+            f"dtype={self.dtype}, "
             f"avail mem={get_available_gpu_memory(self.gpu_id):.2f} GB"
         )
 
@@ -316,7 +318,7 @@ class ModelRunner:
         )
         head_dim = self.model_config.head_dim
         head_num = self.model_config.get_num_kv_heads(self.tp_size)
-        cell_size = head_num * head_dim * self.model_config.num_hidden_layers * 2 * 2
+        cell_size = head_num * head_dim * self.model_config.num_hidden_layers * 2 * torch._utils._element_size(self.dtype)
         rest_memory = available_gpu_memory - total_gpu_memory * (
             1 - self.mem_fraction_static
         )
@@ -337,7 +339,7 @@ class ModelRunner:
         )
         self.token_to_kv_pool = TokenToKVPool(
             self.max_total_num_tokens,
-            dtype=torch.float16,
+            dtype=self.dtype,
             head_num=self.model_config.get_num_kv_heads(self.tp_size),
             head_dim=self.model_config.head_dim,
             layer_num=self.model_config.num_hidden_layers,

python/sglang/srt/managers/controller/tp_worker.py

@@ -120,7 +120,7 @@ class ModelTpServer:
             f"[gpu_id={self.gpu_id}] "
             f"max_total_num_tokens={self.max_total_num_tokens}, "
             f"max_prefill_tokens={self.max_prefill_tokens}, "
-            f"context_len={self.model_config.context_len}, "
+            f"context_len={self.model_config.context_len}"
         )
         if self.tp_rank == 0:
             logger.info(

python/sglang/srt/server_args.py

@@ -11,12 +11,13 @@ class ServerArgs:
     # Model and tokenizer
     model_path: str
     tokenizer_path: Optional[str] = None
-    load_format: str = "auto"
     tokenizer_mode: str = "auto"
-    chat_template: Optional[str] = None
+    load_format: str = "auto"
+    dtype: str = "auto"
     trust_remote_code: bool = True
     context_length: Optional[int] = None
     quantization: Optional[str] = None
+    chat_template: Optional[str] = None
 
     # Port
     host: str = "127.0.0.1"
@@ -107,6 +108,15 @@ class ServerArgs:
             default=[],
             help="The additional ports specified for the server.",
         )
+        parser.add_argument(
+            "--tokenizer-mode",
+            type=str,
+            default=ServerArgs.tokenizer_mode,
+            choices=["auto", "slow"],
+            help="Tokenizer mode. 'auto' will use the fast "
+            "tokenizer if available, and 'slow' will "
+            "always use the slow tokenizer.",
+        )
         parser.add_argument(
             "--load-format",
             type=str,
@@ -124,20 +134,20 @@ class ServerArgs:
             "which is mainly for profiling.",
         )
         parser.add_argument(
-            "--tokenizer-mode",
-            type=str,
-            default=ServerArgs.tokenizer_mode,
-            choices=["auto", "slow"],
-            help="Tokenizer mode. 'auto' will use the fast "
-            "tokenizer if available, and 'slow' will "
-            "always use the slow tokenizer.",
-        )
-        parser.add_argument(
-            "--chat-template",
+            "--dtype",
             type=str,
-            default=ServerArgs.chat_template,
-            help="The buliltin chat template name or the path of the chat template file. This is only used for OpenAI-compatible API server",
-        )
+            default=ServerArgs.dtype,
+            choices=[
+                "auto", "half", "float16", "bfloat16", "float", "float32"
+            ],
+            help='Data type for model weights and activations.\n\n'
+            '* "auto" will use FP16 precision for FP32 and FP16 models, and '
+            'BF16 precision for BF16 models.\n'
+            '* "half" for FP16. Recommended for AWQ quantization.\n'
+            '* "float16" is the same as "half".\n'
+            '* "bfloat16" for a balance between precision and range.\n'
+            '* "float" is shorthand for FP32 precision.\n'
+            '* "float32" for FP32 precision.')
         parser.add_argument(
             "--trust-remote-code",
             action="store_true",
@@ -155,6 +165,12 @@ class ServerArgs:
             default=ServerArgs.quantization,
             help="The quantization method.",
         )
+        parser.add_argument(
+            "--chat-template",
+            type=str,
+            default=ServerArgs.chat_template,
+            help="The buliltin chat template name or the path of the chat template file. This is only used for OpenAI-compatible API server.",
+        )
         parser.add_argument(
             "--mem-fraction-static",
             type=float,