server_args.py 54.3 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
# Copyright 2023-2024 SGLang Team
# 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.
# ==============================================================================
Lianmin Zheng's avatar
Lianmin Zheng committed
14
15
"""The arguments of the server."""

Lianmin Zheng's avatar
Lianmin Zheng committed
16
17
import argparse
import dataclasses
18
import json
19
import logging
20
import os
21
import random
22
import tempfile
23
from typing import List, Literal, Optional
Lianmin Zheng's avatar
Lianmin Zheng committed
24

25
from sglang.srt.hf_transformers_utils import check_gguf_file
Xihuai Wang's avatar
Xihuai Wang committed
26
from sglang.srt.reasoning_parser import ReasoningParser
27
from sglang.srt.utils import (
Vincent's avatar
Vincent committed
28
    configure_ipv6,
HAI's avatar
HAI committed
29
    get_amdgpu_memory_capacity,
30
    get_device,
31
    get_hpu_memory_capacity,
HAI's avatar
HAI committed
32
    get_nvgpu_memory_capacity,
33
    is_cuda,
34
    is_flashinfer_available,
HAI's avatar
HAI committed
35
    is_hip,
36
    is_port_available,
37
    is_remote_url,
38
    is_valid_ipv6_address,
bjmsong's avatar
bjmsong committed
39
    nullable_str,
40
)
41

42
43
logger = logging.getLogger(__name__)

Lianmin Zheng's avatar
Lianmin Zheng committed
44
45
46

@dataclasses.dataclass
class ServerArgs:
Lianmin Zheng's avatar
Lianmin Zheng committed
47
    # Model and tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
48
49
50
    model_path: str
    tokenizer_path: Optional[str] = None
    tokenizer_mode: str = "auto"
51
    skip_tokenizer_init: bool = False
52
    enable_tokenizer_batch_encode: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
53
    load_format: str = "auto"
54
    trust_remote_code: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
55
    dtype: str = "auto"
56
    kv_cache_dtype: str = "auto"
Lianmin Zheng's avatar
Lianmin Zheng committed
57
    quantization: Optional[str] = None
Vincent's avatar
Vincent committed
58
    quantization_param_path: Optional[str] = None
59
    context_length: Optional[int] = None
60
    device: Optional[str] = None
61
    served_model_name: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
62
    chat_template: Optional[str] = None
63
    completion_template: Optional[str] = None
64
    is_embedding: bool = False
65
    revision: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
66

67
    # Port for the HTTP server
Lianmin Zheng's avatar
Lianmin Zheng committed
68
69
70
71
    host: str = "127.0.0.1"
    port: int = 30000

    # Memory and scheduling
Lianmin Zheng's avatar
Lianmin Zheng committed
72
    mem_fraction_static: Optional[float] = None
73
    max_running_requests: Optional[int] = None
74
    max_total_tokens: Optional[int] = None
75
    chunked_prefill_size: Optional[int] = None
76
    max_prefill_tokens: int = 16384
77
    schedule_policy: str = "fcfs"
78
    schedule_conservativeness: float = 1.0
79
    cpu_offload_gb: int = 0
80
    page_size: int = 1
Lianmin Zheng's avatar
Lianmin Zheng committed
81
82
83

    # Other runtime options
    tp_size: int = 1
84
    stream_interval: int = 1
85
    stream_output: bool = False
86
    random_seed: Optional[int] = None
87
    constrained_json_whitespace_pattern: Optional[str] = None
88
    watchdog_timeout: float = 300
89
    dist_timeout: Optional[int] = None  # timeout for torch.distributed
90
    download_dir: Optional[str] = None
91
    base_gpu_id: int = 0
92
    gpu_id_step: int = 1
Lianmin Zheng's avatar
Lianmin Zheng committed
93
94
95

    # Logging
    log_level: str = "info"
96
    log_level_http: Optional[str] = None
97
    log_requests: bool = False
98
    log_requests_level: int = 0
Liangsheng Yin's avatar
Liangsheng Yin committed
99
    show_time_cost: bool = False
100
    enable_metrics: bool = False
101
    decode_log_interval: int = 40
Liangsheng Yin's avatar
Liangsheng Yin committed
102

103
    # API related
104
    api_key: Optional[str] = None
105
    file_storage_path: str = "sglang_storage"
106
    enable_cache_report: bool = False
Xihuai Wang's avatar
Xihuai Wang committed
107
    reasoning_parser: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
108

109
110
111
    # Data parallelism
    dp_size: int = 1
    load_balance_method: str = "round_robin"
112

xiaobochen's avatar
xiaobochen committed
113
114
    # Expert parallelism
    ep_size: int = 1
115

116
    # Multi-node distributed serving
117
    dist_init_addr: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
118
    nnodes: int = 1
119
    node_rank: int = 0
Lianmin Zheng's avatar
Lianmin Zheng committed
120
121
122
123

    # Model override args in JSON
    json_model_override_args: str = "{}"

124
125
126
    # LoRA
    lora_paths: Optional[List[str]] = None
    max_loras_per_batch: int = 8
127
    lora_backend: str = "triton"
128
129

    # Kernel backend
130
131
    attention_backend: Optional[str] = None
    sampling_backend: Optional[str] = None
132
    grammar_backend: Optional[str] = None
133

134
135
    # Speculative decoding
    speculative_algorithm: Optional[str] = None
136
    speculative_draft_model_path: Optional[str] = None
137
138
139
    speculative_num_steps: Optional[int] = None
    speculative_eagle_topk: Optional[int] = None
    speculative_num_draft_tokens: Optional[int] = None
140
141
    speculative_accept_threshold_single: float = 1.0
    speculative_accept_threshold_acc: float = 1.0
142
    speculative_token_map: Optional[str] = None
143
144
145

    # Double Sparsity
    enable_double_sparsity: bool = False
Vincent's avatar
Vincent committed
146
    ds_channel_config_path: Optional[str] = None
147
148
149
150
151
    ds_heavy_channel_num: int = 32
    ds_heavy_token_num: int = 256
    ds_heavy_channel_type: str = "qk"
    ds_sparse_decode_threshold: int = 4096

152
    # Optimization/debug options
Lianmin Zheng's avatar
Lianmin Zheng committed
153
    disable_radix_cache: bool = False
154
    disable_cuda_graph: bool = False
155
    disable_cuda_graph_padding: bool = False
156
    enable_nccl_nvls: bool = False
157
    disable_outlines_disk_cache: bool = False
158
    disable_custom_all_reduce: bool = False
159
    enable_llama4_multimodal: Optional[bool] = None
160
    disable_overlap_schedule: bool = False
161
    enable_mixed_chunk: bool = False
Ke Bao's avatar
Ke Bao committed
162
    enable_dp_attention: bool = False
xiaobochen's avatar
xiaobochen committed
163
    enable_ep_moe: bool = False
164
    enable_deepep_moe: bool = False
165
    deepep_mode: Optional[Literal["auto", "normal", "low_latency"]] = "auto"
166
    enable_torch_compile: bool = False
167
    torch_compile_max_bs: int = 32
168
    cuda_graph_max_bs: Optional[int] = None
169
    cuda_graph_bs: Optional[List[int]] = None
170
    torchao_config: str = ""
171
    enable_nan_detection: bool = False
172
    enable_p2p_check: bool = False
173
    triton_attention_reduce_in_fp32: bool = False
174
    triton_attention_num_kv_splits: int = 8
175
    num_continuous_decode_steps: int = 1
176
    delete_ckpt_after_loading: bool = False
177
    enable_memory_saver: bool = False
178
    allow_auto_truncate: bool = False
179
    enable_custom_logit_processor: bool = False
Vincent's avatar
Vincent committed
180
    tool_call_parser: Optional[str] = None
181
    enable_hierarchical_cache: bool = False
182
    hicache_ratio: float = 2.0
Zhiqiang Xie's avatar
Zhiqiang Xie committed
183
184
    hicache_size: int = 0
    hicache_write_policy: str = "write_through_selective"
185
    flashinfer_mla_disable_ragged: bool = False
186
    warmups: Optional[str] = None
187
    moe_dense_tp_size: Optional[int] = None
188
    n_share_experts_fusion: int = 0
189
    disable_chunked_prefix_cache: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
190
    disable_fast_image_processor: bool = False
191
192
193
194
195

    # Debug tensor dumps
    debug_tensor_dump_output_folder: Optional[str] = None
    debug_tensor_dump_input_file: Optional[str] = None
    debug_tensor_dump_inject: bool = False
196

Byron Hsu's avatar
Byron Hsu committed
197
198
199
    # For PD disaggregation: can be "null" (not disaggregated), "prefill" (prefill-only), or "decode" (decode-only)
    disaggregation_mode: str = "null"
    disaggregation_bootstrap_port: int = 8998
200
    disaggregation_transfer_backend: str = "mooncake"
201
    disaggregation_ib_device: Optional[str] = None
Byron Hsu's avatar
Byron Hsu committed
202

Lianmin Zheng's avatar
Lianmin Zheng committed
203
    def __post_init__(self):
204
205
206
207
208
209
210
        # Expert parallelism
        if self.enable_ep_moe:
            self.ep_size = self.tp_size
            logger.info(
                f"EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
            )

211
        # Set missing default values
Lianmin Zheng's avatar
Lianmin Zheng committed
212
213
        if self.tokenizer_path is None:
            self.tokenizer_path = self.model_path
214

215
216
217
        if self.device is None:
            self.device = get_device()

218
219
220
        if self.served_model_name is None:
            self.served_model_name = self.model_path

221
222
223
        if self.random_seed is None:
            self.random_seed = random.randint(0, 1 << 30)

224
        if is_cuda():
225
            gpu_mem = get_nvgpu_memory_capacity()
226
227
        elif is_hip():
            gpu_mem = get_amdgpu_memory_capacity()
228
229
        elif self.device == "hpu":
            gpu_mem = get_hpu_memory_capacity()
230
231
232
        else:
            # GPU memory is not known yet or no GPU is available.
            gpu_mem = None
233
234

        # Set mem fraction static, which depends on the tensor parallelism size
Lianmin Zheng's avatar
Lianmin Zheng committed
235
        if self.mem_fraction_static is None:
236
            if self.tp_size >= 16:
237
                self.mem_fraction_static = 0.79
238
            elif self.tp_size >= 8:
239
                self.mem_fraction_static = 0.81
Lianmin Zheng's avatar
Lianmin Zheng committed
240
            elif self.tp_size >= 4:
241
                self.mem_fraction_static = 0.85
Lianmin Zheng's avatar
Lianmin Zheng committed
242
            elif self.tp_size >= 2:
243
                self.mem_fraction_static = 0.87
Ying Sheng's avatar
Ying Sheng committed
244
            else:
245
                self.mem_fraction_static = 0.88
246

247
248
        # Set chunked prefill size, which depends on the gpu memory capacity
        if self.chunked_prefill_size is None:
249
            if gpu_mem is not None and gpu_mem < 25_000:
250
251
252
                self.chunked_prefill_size = 2048
            else:
                self.chunked_prefill_size = 8192
253

Lianmin Zheng's avatar
Lianmin Zheng committed
254
255
        assert self.chunked_prefill_size % self.page_size == 0

256
257
258
259
260
        assert self.moe_dense_tp_size in {
            1,
            None,
        }, f"moe_dense_tp_size only support 1 and None currently"

261
        if self.attention_backend == "flashmla":
262
263
264
265
            logger.warning(
                "FlashMLA only supports a page_size of 64, change page_size to 64."
            )
            self.page_size = 64
266
267
        # Set cuda graph max batch size
        if self.cuda_graph_max_bs is None:
268
            # Based on detailed statistics, when serving TP1/TP2 models on lower-end GPUs with HBM<25G, you can either disable cuda graph or set `cuda_graph_max_bs` to a very small value to reduce the memory overhead of creating cuda graphs, with almost no impact on performance. However, when serving models with TP4 or TP8, we need to enable cuda graph to maintain high performance. In this case, we can set `cuda_graph_max_bs` to 80 (half of the default value 160) to reduce the memory overhead of creating cuda graphs. Looking at the logs from TP4 serving of qwen2-72b, a value of 80 is sufficient and can reduce the memory overhead of creating cuda graphs on lower-end GPUs compared to the original 160, avoiding OOM issues.
269
            if gpu_mem is not None and gpu_mem < 25_000:
270
271
272
273
                if self.tp_size < 4:
                    self.cuda_graph_max_bs = 8
                else:
                    self.cuda_graph_max_bs = 80
274
275
            else:
                self.cuda_graph_max_bs = 160
276

277
        # Set kernel backends for hpu device
278
279
280
281
        if self.device == "hpu":
            self.attention_backend = "torch_native"
            self.sampling_backend = "pytorch"

282
        if self.sampling_backend is None:
283
284
285
286
287
            self.sampling_backend = (
                "flashinfer" if is_flashinfer_available() else "pytorch"
            )

        if self.attention_backend == "torch_native":
288
            logger.warning(
289
290
291
                "Cuda graph is disabled because of using torch native attention backend"
            )
            self.disable_cuda_graph = True
292

293
294
295
        # Choose grammar backend
        if self.grammar_backend is None:
            self.grammar_backend = "xgrammar"
296

297
298
299
300
301
302
303
        # Expert parallelism
        if self.enable_ep_moe:
            self.ep_size = self.tp_size
            logger.info(
                f"EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
            )

304
305
        self.enable_multimodal: Optional[bool] = self.enable_llama4_multimodal

306
        # Data parallelism attention
Ke Bao's avatar
Ke Bao committed
307
        if self.enable_dp_attention:
308
            self.schedule_conservativeness = self.schedule_conservativeness * 0.3
Lianmin Zheng's avatar
Lianmin Zheng committed
309
310
311
312
313
            assert (
                self.dp_size > 1
            ), "Please set a dp-size > 1. You can use 1 < dp-size <= tp-size "
            assert self.tp_size % self.dp_size == 0
            self.chunked_prefill_size = self.chunked_prefill_size // self.dp_size
314
            logger.warning(
315
                f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
316
            )
317
318
319
320

        self.enable_sp_layernorm = False
        # DeepEP MoE
        if self.enable_deepep_moe:
321
322
323
324
            if self.deepep_mode == "auto":
                assert (
                    not self.enable_dp_attention
                ), "DeepEP MoE `auto` mode is not supported with DP Attention."
325
326
327
328
329
330
331
            self.ep_size = self.tp_size
            self.enable_sp_layernorm = (
                self.dp_size < self.tp_size if self.enable_dp_attention else True
            )
            logger.info(
                f"DeepEP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
            )
332

333
        # Speculative Decoding
334
335
336
337
        if self.speculative_algorithm == "NEXTN":
            # NEXTN shares the same implementation of EAGLE
            self.speculative_algorithm = "EAGLE"

James Liu's avatar
James Liu committed
338
339
340
341
        if (
            self.speculative_algorithm == "EAGLE"
            or self.speculative_algorithm == "EAGLE3"
        ):
342
            if self.max_running_requests is None:
343
                self.max_running_requests = 48
344
            self.disable_overlap_schedule = True
345
            logger.info(
346
                "Overlap scheduler is disabled because of using "
347
                "eagle speculative decoding."
348
            )
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363

            # Auto choose parameters
            if self.speculative_num_steps is None:
                assert (
                    self.speculative_eagle_topk is None
                    and self.speculative_num_draft_tokens is None
                )
                (
                    self.speculative_num_steps,
                    self.speculative_eagle_topk,
                    self.speculative_num_draft_tokens,
                ) = auto_choose_speculative_params(self)

            if self.page_size > 1 and self.speculative_eagle_topk > 1:
                self.speculative_eagle_topk = 1
364
365
366
367
368
369
370
371
372
373
374
375
                logger.info(
                    "speculative_eagle_topk is adjusted to 1 when page_size > 1"
                )

            if (
                self.speculative_eagle_topk == 1
                and self.speculative_num_draft_tokens != self.speculative_num_steps + 1
            ):
                logger.info(
                    "speculative_num_draft_tokens is adjusted to speculative_num_steps + 1 when speculative_eagle_topk == 1"
                )
                self.speculative_num_draft_tokens = self.speculative_num_steps + 1
376

377
            # The token generated from the verify step is counted.
378
            # If sepculative_num_steps >= speculative_num_draft_tokens, the additional tokens will definitely be discarded.
379
            # assert self.speculative_num_steps < self.speculative_num_draft_tokens
380

381
382
383
384
385
386
        # GGUF
        if (
            self.load_format == "auto" or self.load_format == "gguf"
        ) and check_gguf_file(self.model_path):
            self.quantization = self.load_format = "gguf"

387
388
389
        if is_remote_url(self.model_path):
            self.load_format = "remote"

390
391
392
393
        # AMD-specific Triton attention KV splits default number
        if is_hip():
            self.triton_attention_num_kv_splits = 16

Byron Hsu's avatar
Byron Hsu committed
394
395
396
397
398
399
400
401
402
403
404
405
        # PD disaggregation
        if self.disaggregation_mode == "prefill":
            self.disable_cuda_graph = True
            logger.warning("KV cache is forced as chunk cache for decode server")
            self.disable_overlap_schedule = True
            logger.warning("Overlap scheduler is disabled for prefill server")
        elif self.disaggregation_mode == "decode":
            self.disable_radix_cache = True
            logger.warning("Cuda graph is disabled for prefill server")
            self.disable_overlap_schedule = True
            logger.warning("Overlap scheduler is disabled for decode server")

406
407
408
        os.environ["SGLANG_ENABLE_TORCH_COMPILE"] = (
            "1" if self.enable_torch_compile else "0"
        )
409
410
411
412
        # Set env var before grammar backends init
        os.environ["SGLANG_DISABLE_OUTLINES_DISK_CACHE"] = (
            "1" if self.disable_outlines_disk_cache else "0"
        )
413

Lianmin Zheng's avatar
Lianmin Zheng committed
414
415
    @staticmethod
    def add_cli_args(parser: argparse.ArgumentParser):
416
        # Model and port args
Lianmin Zheng's avatar
Lianmin Zheng committed
417
418
419
420
421
422
423
424
425
426
427
428
        parser.add_argument(
            "--model-path",
            type=str,
            help="The path of the model weights. This can be a local folder or a Hugging Face repo ID.",
            required=True,
        )
        parser.add_argument(
            "--tokenizer-path",
            type=str,
            default=ServerArgs.tokenizer_path,
            help="The path of the tokenizer.",
        )
Yuanhan Zhang's avatar
Yuanhan Zhang committed
429
430
431
432
433
434
        parser.add_argument(
            "--host", type=str, default=ServerArgs.host, help="The host of the server."
        )
        parser.add_argument(
            "--port", type=int, default=ServerArgs.port, help="The port of the server."
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
435
436
437
438
439
440
441
442
443
        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.",
        )
444
445
446
447
448
        parser.add_argument(
            "--skip-tokenizer-init",
            action="store_true",
            help="If set, skip init tokenizer and pass input_ids in generate request",
        )
449
450
451
452
453
        parser.add_argument(
            "--enable-tokenizer-batch-encode",
            action="store_true",
            help="Enable batch tokenization for improved performance when processing multiple text inputs. Do not use with image inputs, pre-tokenized input_ids, or input_embeds.",
        )
454
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
455
456
457
            "--load-format",
            type=str,
            default=ServerArgs.load_format,
458
459
460
461
462
463
            choices=[
                "auto",
                "pt",
                "safetensors",
                "npcache",
                "dummy",
464
                "sharded_state",
465
466
                "gguf",
                "bitsandbytes",
467
                "layered",
468
                "remote",
469
            ],
Lianmin Zheng's avatar
Lianmin Zheng committed
470
471
472
473
474
475
476
477
478
            help="The format of the model weights to load. "
            '"auto" will try to load the weights in the safetensors format '
            "and fall back to the pytorch bin format if safetensors format "
            "is not available. "
            '"pt" will load the weights in the pytorch bin format. '
            '"safetensors" will load the weights in the safetensors format. '
            '"npcache" will load the weights in pytorch format and store '
            "a numpy cache to speed up the loading. "
            '"dummy" will initialize the weights with random values, '
479
            "which is mainly for profiling."
480
481
            '"gguf" will load the weights in the gguf format. '
            '"bitsandbytes" will load the weights using bitsandbytes '
482
483
484
485
            "quantization."
            '"layered" loads weights layer by layer so that one can quantize a '
            "layer before loading another to make the peak memory envelope "
            "smaller.",
Lianmin Zheng's avatar
Lianmin Zheng committed
486
        )
487
488
489
490
491
        parser.add_argument(
            "--trust-remote-code",
            action="store_true",
            help="Whether or not to allow for custom models defined on the Hub in their own modeling files.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
492
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
493
            "--dtype",
Cody Yu's avatar
Cody Yu committed
494
            type=str,
Lianmin Zheng's avatar
Lianmin Zheng committed
495
            default=ServerArgs.dtype,
Ying Sheng's avatar
Ying Sheng committed
496
497
            choices=["auto", "half", "float16", "bfloat16", "float", "float32"],
            help="Data type for model weights and activations.\n\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
498
            '* "auto" will use FP16 precision for FP32 and FP16 models, and '
Ying Sheng's avatar
Ying Sheng committed
499
            "BF16 precision for BF16 models.\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
500
501
502
503
            '* "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'
Ying Sheng's avatar
Ying Sheng committed
504
505
            '* "float32" for FP32 precision.',
        )
506
507
508
509
        parser.add_argument(
            "--kv-cache-dtype",
            type=str,
            default=ServerArgs.kv_cache_dtype,
bjmsong's avatar
bjmsong committed
510
511
512
            choices=["auto", "fp8_e5m2", "fp8_e4m3"],
            help='Data type for kv cache storage. "auto" will use model data type. "fp8_e5m2" and "fp8_e4m3" is supported for CUDA 11.8+.',
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
513
514
515
516
        parser.add_argument(
            "--quantization",
            type=str,
            default=ServerArgs.quantization,
Ying Sheng's avatar
Ying Sheng committed
517
518
519
520
521
522
            choices=[
                "awq",
                "fp8",
                "gptq",
                "marlin",
                "gptq_marlin",
Ying Sheng's avatar
Ying Sheng committed
523
                "awq_marlin",
Ying Sheng's avatar
Ying Sheng committed
524
                "bitsandbytes",
525
                "gguf",
526
                "modelopt",
527
                "modelopt_fp4",
528
                "w8a8_int8",
HandH1998's avatar
HandH1998 committed
529
                "w8a8_fp8",
AniZpZ's avatar
AniZpZ committed
530
                "moe_wna16",
Ying Sheng's avatar
Ying Sheng committed
531
            ],
Lianmin Zheng's avatar
Lianmin Zheng committed
532
533
            help="The quantization method.",
        )
534
535
536
537
538
539
540
541
542
        parser.add_argument(
            "--quantization-param-path",
            type=nullable_str,
            default=None,
            help="Path to the JSON file containing the KV cache "
            "scaling factors. This should generally be supplied, when "
            "KV cache dtype is FP8. Otherwise, KV cache scaling factors "
            "default to 1.0, which may cause accuracy issues. ",
        )
543
544
545
546
547
548
549
550
551
        parser.add_argument(
            "--context-length",
            type=int,
            default=ServerArgs.context_length,
            help="The model's maximum context length. Defaults to None (will use the value from the model's config.json instead).",
        )
        parser.add_argument(
            "--device",
            type=str,
552
553
            default=ServerArgs.device,
            help="The device to use ('cuda', 'xpu', 'hpu', 'cpu'). Defaults to auto-detection if not specified.",
554
        )
555
556
557
558
559
560
        parser.add_argument(
            "--served-model-name",
            type=str,
            default=ServerArgs.served_model_name,
            help="Override the model name returned by the v1/models endpoint in OpenAI API server.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
561
562
563
564
565
566
        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.",
        )
567
568
569
570
571
572
        parser.add_argument(
            "--completion-template",
            type=str,
            default=ServerArgs.completion_template,
            help="The buliltin completion template name or the path of the completion template file. This is only used for OpenAI-compatible API server. only for code completion currently.",
        )
573
574
575
576
577
        parser.add_argument(
            "--is-embedding",
            action="store_true",
            help="Whether to use a CausalLM as an embedding model.",
        )
578
579
580
581
582
583
584
585
        parser.add_argument(
            "--revision",
            type=str,
            default=None,
            help="The specific model version to use. It can be a branch "
            "name, a tag name, or a commit id. If unspecified, will use "
            "the default version.",
        )
586
        # Memory and scheduling
Lianmin Zheng's avatar
Lianmin Zheng committed
587
588
589
590
        parser.add_argument(
            "--mem-fraction-static",
            type=float,
            default=ServerArgs.mem_fraction_static,
591
            help="The fraction of the memory used for static allocation (model weights and KV cache memory pool). Use a smaller value if you see out-of-memory errors.",
Lianmin Zheng's avatar
Lianmin Zheng committed
592
        )
593
594
595
596
597
598
        parser.add_argument(
            "--max-running-requests",
            type=int,
            default=ServerArgs.max_running_requests,
            help="The maximum number of running requests.",
        )
599
600
601
602
        parser.add_argument(
            "--max-total-tokens",
            type=int,
            default=ServerArgs.max_total_tokens,
603
604
            help="The maximum number of tokens in the memory pool. If not specified, it will be automatically calculated based on the memory usage fraction. "
            "This option is typically used for development and debugging purposes.",
605
        )
606
607
608
609
        parser.add_argument(
            "--chunked-prefill-size",
            type=int,
            default=ServerArgs.chunked_prefill_size,
610
            help="The maximum number of tokens in a chunk for the chunked prefill. Setting this to -1 means disabling chunked prefill.",
611
612
613
614
615
616
617
        )
        parser.add_argument(
            "--max-prefill-tokens",
            type=int,
            default=ServerArgs.max_prefill_tokens,
            help="The maximum number of tokens in a prefill batch. The real bound will be the maximum of this value and the model's maximum context length.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
618
        parser.add_argument(
619
            "--schedule-policy",
Lianmin Zheng's avatar
Lianmin Zheng committed
620
            type=str,
621
            default=ServerArgs.schedule_policy,
Liangsheng Yin's avatar
Liangsheng Yin committed
622
            choices=["lpm", "random", "fcfs", "dfs-weight"],
623
            help="The scheduling policy of the requests.",
Lianmin Zheng's avatar
Lianmin Zheng committed
624
        )
625
626
627
628
        parser.add_argument(
            "--schedule-conservativeness",
            type=float,
            default=ServerArgs.schedule_conservativeness,
629
            help="How conservative the schedule policy is. A larger value means more conservative scheduling. Use a larger value if you see requests being retracted frequently.",
630
        )
631
632
633
634
        parser.add_argument(
            "--cpu-offload-gb",
            type=int,
            default=ServerArgs.cpu_offload_gb,
635
            help="How many GBs of RAM to reserve for CPU offloading.",
636
        )
637
638
639
640
641
642
        parser.add_argument(
            "--page-size",
            type=int,
            default=ServerArgs.page_size,
            help="The number of tokens in a page.",
        )
643

644
        # Other runtime options
Lianmin Zheng's avatar
Lianmin Zheng committed
645
        parser.add_argument(
646
            "--tensor-parallel-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
647
            "--tp-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
648
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
649
            default=ServerArgs.tp_size,
650
            help="The tensor parallelism size.",
651
        )
652
653
654
        parser.add_argument(
            "--stream-interval",
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
655
            default=ServerArgs.stream_interval,
656
            help="The interval (or buffer size) for streaming in terms of the token length. A smaller value makes streaming smoother, while a larger value makes the throughput higher",
657
        )
658
659
660
661
662
        parser.add_argument(
            "--stream-output",
            action="store_true",
            help="Whether to output as a sequence of disjoint segments.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
663
664
665
666
        parser.add_argument(
            "--random-seed",
            type=int,
            default=ServerArgs.random_seed,
667
            help="The random seed.",
Lianmin Zheng's avatar
Lianmin Zheng committed
668
        )
669
670
671
672
673
674
        parser.add_argument(
            "--constrained-json-whitespace-pattern",
            type=str,
            default=ServerArgs.constrained_json_whitespace_pattern,
            help=r"Regex pattern for syntactic whitespaces allowed in JSON constrained output. For example, to allow the model generate consecutive whitespaces, set the pattern to [\n\t ]*",
        )
675
676
677
678
679
680
        parser.add_argument(
            "--watchdog-timeout",
            type=float,
            default=ServerArgs.watchdog_timeout,
            help="Set watchdog timeout in seconds. If a forward batch takes longer than this, the server will crash to prevent hanging.",
        )
681
682
683
684
685
686
        parser.add_argument(
            "--dist-timeout",
            type=int,
            default=ServerArgs.dist_timeout,
            help="Set timeout for torch.distributed initialization.",
        )
687
688
689
690
        parser.add_argument(
            "--download-dir",
            type=str,
            default=ServerArgs.download_dir,
Lianmin Zheng's avatar
Lianmin Zheng committed
691
            help="Model download directory for huggingface.",
692
        )
693
694
695
696
697
698
        parser.add_argument(
            "--base-gpu-id",
            type=int,
            default=ServerArgs.base_gpu_id,
            help="The base GPU ID to start allocating GPUs from. Useful when running multiple instances on the same machine.",
        )
699
700
701
702
703
704
        parser.add_argument(
            "--gpu-id-step",
            type=int,
            default=ServerArgs.gpu_id_step,
            help="The delta between consecutive GPU IDs that are used. For example, setting it to 2 will use GPU 0,2,4,...",
        )
705
706

        # Logging
Lianmin Zheng's avatar
Lianmin Zheng committed
707
708
709
710
        parser.add_argument(
            "--log-level",
            type=str,
            default=ServerArgs.log_level,
711
            help="The logging level of all loggers.",
Lianmin Zheng's avatar
Lianmin Zheng committed
712
        )
713
        parser.add_argument(
714
715
716
717
            "--log-level-http",
            type=str,
            default=ServerArgs.log_level_http,
            help="The logging level of HTTP server. If not set, reuse --log-level by default.",
718
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
719
        parser.add_argument(
720
            "--log-requests",
Lianmin Zheng's avatar
Lianmin Zheng committed
721
            action="store_true",
722
723
724
725
726
727
728
729
            help="Log metadata, inputs, outputs of all requests. The verbosity is decided by --log-requests-level",
        )
        parser.add_argument(
            "--log-requests-level",
            type=int,
            default=0,
            help="0: Log metadata. 1. Log metadata and partial input/output. 2. Log every input/output.",
            choices=[0, 1, 2],
Lianmin Zheng's avatar
Lianmin Zheng committed
730
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
731
732
733
        parser.add_argument(
            "--show-time-cost",
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
734
            help="Show time cost of custom marks.",
Lianmin Zheng's avatar
Lianmin Zheng committed
735
        )
736
737
738
739
740
        parser.add_argument(
            "--enable-metrics",
            action="store_true",
            help="Enable log prometheus metrics.",
        )
741
742
743
744
        parser.add_argument(
            "--decode-log-interval",
            type=int,
            default=ServerArgs.decode_log_interval,
745
            help="The log interval of decode batch.",
746
        )
747

748
        # API related
Liangsheng Yin's avatar
Liangsheng Yin committed
749
750
751
752
        parser.add_argument(
            "--api-key",
            type=str,
            default=ServerArgs.api_key,
753
            help="Set API key of the server. It is also used in the OpenAI API compatible server.",
Liangsheng Yin's avatar
Liangsheng Yin committed
754
        )
755
        parser.add_argument(
756
            "--file-storage-path",
757
            type=str,
758
            default=ServerArgs.file_storage_path,
759
760
            help="The path of the file storage in backend.",
        )
761
762
763
764
765
        parser.add_argument(
            "--enable-cache-report",
            action="store_true",
            help="Return number of cached tokens in usage.prompt_tokens_details for each openai request.",
        )
Xihuai Wang's avatar
Xihuai Wang committed
766
767
768
769
770
771
772
        parser.add_argument(
            "--reasoning-parser",
            type=str,
            choices=list(ReasoningParser.DetectorMap.keys()),
            default=ServerArgs.reasoning_parser,
            help=f"Specify the parser for reasoning models, supported parsers are: {list(ReasoningParser.DetectorMap.keys())}.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
773

774
775
        # Data parallelism
        parser.add_argument(
776
            "--data-parallel-size",
777
778
779
            "--dp-size",
            type=int,
            default=ServerArgs.dp_size,
780
            help="The data parallelism size.",
781
782
783
784
785
        )
        parser.add_argument(
            "--load-balance-method",
            type=str,
            default=ServerArgs.load_balance_method,
786
            help="The load balancing strategy for data parallelism.",
787
788
789
790
791
            choices=[
                "round_robin",
                "shortest_queue",
            ],
        )
792

xiaobochen's avatar
xiaobochen committed
793
794
795
796
797
798
799
800
        # Expert parallelism
        parser.add_argument(
            "--expert-parallel-size",
            "--ep-size",
            type=int,
            default=ServerArgs.ep_size,
            help="The expert parallelism size.",
        )
801

802
        # Multi-node distributed serving
803
        parser.add_argument(
804
805
            "--dist-init-addr",
            "--nccl-init-addr",  # For backward compatbility. This will be removed in the future.
806
            type=str,
807
            help="The host address for initializing distributed backend (e.g., `192.168.0.2:25000`).",
808
809
        )
        parser.add_argument(
Liangsheng Yin's avatar
Liangsheng Yin committed
810
            "--nnodes", type=int, default=ServerArgs.nnodes, help="The number of nodes."
811
        )
812
813
814
        parser.add_argument(
            "--node-rank", type=int, default=ServerArgs.node_rank, help="The node rank."
        )
815

Lianmin Zheng's avatar
Lianmin Zheng committed
816
817
818
819
820
821
822
823
        # Model override args
        parser.add_argument(
            "--json-model-override-args",
            type=str,
            help="A dictionary in JSON string format used to override default model configurations.",
            default=ServerArgs.json_model_override_args,
        )

824
825
826
827
828
829
830
        # LoRA
        parser.add_argument(
            "--lora-paths",
            type=str,
            nargs="*",
            default=None,
            action=LoRAPathAction,
831
            help="The list of LoRA adapters. You can provide a list of either path in str or renamed path in the format {name}={path}.",
832
833
834
835
836
        )
        parser.add_argument(
            "--max-loras-per-batch",
            type=int,
            default=8,
837
838
839
840
841
842
843
            help="Maximum number of adapters for a running batch, include base-only request.",
        )
        parser.add_argument(
            "--lora-backend",
            type=str,
            default="triton",
            help="Choose the kernel backend for multi-LoRA serving.",
844
845
846
        )

        # Kernel backend
847
848
849
        parser.add_argument(
            "--attention-backend",
            type=str,
850
            choices=["flashinfer", "triton", "torch_native", "fa3", "flashmla"],
851
852
853
            default=ServerArgs.attention_backend,
            help="Choose the kernels for attention layers.",
        )
854
855
856
857
858
859
860
        parser.add_argument(
            "--sampling-backend",
            type=str,
            choices=["flashinfer", "pytorch"],
            default=ServerArgs.sampling_backend,
            help="Choose the kernels for sampling layers.",
        )
861
862
863
        parser.add_argument(
            "--grammar-backend",
            type=str,
864
            choices=["xgrammar", "outlines", "llguidance", "none"],
865
            default=ServerArgs.grammar_backend,
Lianmin Zheng's avatar
Lianmin Zheng committed
866
            help="Choose the backend for grammar-guided decoding.",
867
        )
868
869
        parser.add_argument(
            "--enable-flashinfer-mla",
870
871
            action=DeprecatedAction,
            help="--enable-flashinfer-mla is deprecated. Please use '--attention-backend flashinfer' instead.",
872
        )
lukec's avatar
lukec committed
873
874
        parser.add_argument(
            "--enable-flashmla",
875
876
            action=DeprecatedAction,
            help="--enable-flashmla is deprecated. Please use '--attention-backend flashmla' instead.",
lukec's avatar
lukec committed
877
        )
878
879
880
881
882
        parser.add_argument(
            "--flashinfer-mla-disable-ragged",
            action="store_true",
            help="Not using ragged prefill wrapper when running flashinfer mla",
        )
883

884
885
886
887
        # Speculative decoding
        parser.add_argument(
            "--speculative-algorithm",
            type=str,
James Liu's avatar
James Liu committed
888
            choices=["EAGLE", "EAGLE3", "NEXTN"],
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
            help="Speculative algorithm.",
        )
        parser.add_argument(
            "--speculative-draft-model-path",
            type=str,
            help="The path of the draft model weights. This can be a local folder or a Hugging Face repo ID.",
        )
        parser.add_argument(
            "--speculative-num-steps",
            type=int,
            help="The number of steps sampled from draft model in Speculative Decoding.",
            default=ServerArgs.speculative_num_steps,
        )
        parser.add_argument(
            "--speculative-eagle-topk",
            type=int,
905
            help="The number of tokens sampled from the draft model in eagle2 each step.",
906
907
            default=ServerArgs.speculative_eagle_topk,
        )
908
909
910
        parser.add_argument(
            "--speculative-num-draft-tokens",
            type=int,
911
            help="The number of tokens sampled from the draft model in Speculative Decoding.",
912
913
            default=ServerArgs.speculative_num_draft_tokens,
        )
914
915
916
917
918
919
920
921
922
923
924
925
        parser.add_argument(
            "--speculative-accept-threshold-single",
            type=float,
            help="Accept a draft token if its probability in the target model is greater than this threshold.",
            default=ServerArgs.speculative_accept_threshold_single,
        )
        parser.add_argument(
            "--speculative-accept-threshold-acc",
            type=float,
            help="The accept probability of a draft token is raised from its target probability p to min(1, p / threshold_acc).",
            default=ServerArgs.speculative_accept_threshold_acc,
        )
926
927
928
929
930
931
        parser.add_argument(
            "--speculative-token-map",
            type=str,
            help="The path of the draft model's small vocab table.",
            default=ServerArgs.speculative_token_map,
        )
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969

        # Double Sparsity
        parser.add_argument(
            "--enable-double-sparsity",
            action="store_true",
            help="Enable double sparsity attention",
        )
        parser.add_argument(
            "--ds-channel-config-path",
            type=str,
            default=ServerArgs.ds_channel_config_path,
            help="The path of the double sparsity channel config",
        )
        parser.add_argument(
            "--ds-heavy-channel-num",
            type=int,
            default=ServerArgs.ds_heavy_channel_num,
            help="The number of heavy channels in double sparsity attention",
        )
        parser.add_argument(
            "--ds-heavy-token-num",
            type=int,
            default=ServerArgs.ds_heavy_token_num,
            help="The number of heavy tokens in double sparsity attention",
        )
        parser.add_argument(
            "--ds-heavy-channel-type",
            type=str,
            default=ServerArgs.ds_heavy_channel_type,
            help="The type of heavy channels in double sparsity attention",
        )
        parser.add_argument(
            "--ds-sparse-decode-threshold",
            type=int,
            default=ServerArgs.ds_sparse_decode_threshold,
            help="The type of heavy channels in double sparsity attention",
        )

970
        # Optimization/debug options
Liangsheng Yin's avatar
Liangsheng Yin committed
971
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
972
            "--disable-radix-cache",
Liangsheng Yin's avatar
Liangsheng Yin committed
973
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
974
            help="Disable RadixAttention for prefix caching.",
Liangsheng Yin's avatar
Liangsheng Yin committed
975
        )
976
977
978
979
980
        parser.add_argument(
            "--disable-cuda-graph",
            action="store_true",
            help="Disable cuda graph.",
        )
981
        parser.add_argument(
982
983
984
985
            "--disable-cuda-graph-padding",
            action="store_true",
            help="Disable cuda graph when padding is needed. Still uses cuda graph when padding is not needed.",
        )
986
987
988
989
990
        parser.add_argument(
            "--enable-nccl-nvls",
            action="store_true",
            help="Enable NCCL NVLS for prefill heavy requests when available.",
        )
991
        parser.add_argument(
992
            "--disable-outlines-disk-cache",
993
            action="store_true",
994
            help="Disable disk cache of outlines to avoid possible crashes related to file system or high concurrency.",
995
        )
996
997
998
999
1000
        parser.add_argument(
            "--disable-custom-all-reduce",
            action="store_true",
            help="Disable the custom all-reduce kernel and fall back to NCCL.",
        )
1001
1002
        parser.add_argument(
            "--enable-llama4-multimodal",
1003
            default=ServerArgs.enable_llama4_multimodal,
1004
1005
1006
            action="store_true",
            help="Enable the multimodal functionality for Llama-4.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1007
        parser.add_argument(
1008
            "--disable-overlap-schedule",
Lianmin Zheng's avatar
Lianmin Zheng committed
1009
            action="store_true",
1010
            help="Disable the overlap scheduler, which overlaps the CPU scheduler with GPU model worker.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1011
        )
1012
1013
1014
        parser.add_argument(
            "--enable-mixed-chunk",
            action="store_true",
1015
            help="Enabling mixing prefill and decode in a batch when using chunked prefill.",
1016
        )
Ke Bao's avatar
Ke Bao committed
1017
1018
1019
1020
1021
        parser.add_argument(
            "--enable-dp-attention",
            action="store_true",
            help="Enabling data parallelism for attention and tensor parallelism for FFN. The dp size should be equal to the tp size. Currently only DeepSeek-V2 is supported.",
        )
xiaobochen's avatar
xiaobochen committed
1022
1023
1024
1025
1026
        parser.add_argument(
            "--enable-ep-moe",
            action="store_true",
            help="Enabling expert parallelism for moe. The ep size is equal to the tp size.",
        )
1027
1028
1029
        parser.add_argument(
            "--enable-torch-compile",
            action="store_true",
1030
1031
            help="Optimize the model with torch.compile. Experimental feature.",
        )
1032
        parser.add_argument(
1033
            "--torch-compile-max-bs",
1034
            type=int,
1035
            default=ServerArgs.torch_compile_max_bs,
1036
1037
            help="Set the maximum batch size when using torch compile.",
        )
1038
        parser.add_argument(
1039
            "--cuda-graph-max-bs",
1040
            type=int,
1041
            default=ServerArgs.cuda_graph_max_bs,
1042
1043
            help="Set the maximum batch size for cuda graph.",
        )
1044
1045
1046
1047
1048
1049
        parser.add_argument(
            "--cuda-graph-bs",
            type=int,
            nargs="+",
            help="Set the list of batch sizes for cuda graph.",
        )
1050
1051
1052
1053
        parser.add_argument(
            "--torchao-config",
            type=str,
            default=ServerArgs.torchao_config,
1054
            help="Optimize the model with torchao. Experimental feature. Current choices are: int8dq, int8wo, int4wo-<group_size>, fp8wo, fp8dq-per_tensor, fp8dq-per_row",
1055
        )
1056
1057
1058
1059
1060
        parser.add_argument(
            "--enable-nan-detection",
            action="store_true",
            help="Enable the NaN detection for debugging purposes.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1061
        parser.add_argument(
1062
            "--enable-p2p-check",
Lianmin Zheng's avatar
Lianmin Zheng committed
1063
            action="store_true",
1064
            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1065
        )
1066
        parser.add_argument(
1067
            "--triton-attention-reduce-in-fp32",
1068
            action="store_true",
1069
            help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16."
1070
            "This only affects Triton attention kernels.",
1071
        )
1072
1073
1074
1075
1076
1077
        parser.add_argument(
            "--triton-attention-num-kv-splits",
            type=int,
            default=ServerArgs.triton_attention_num_kv_splits,
            help="The number of KV splits in flash decoding Triton kernel. Larger value is better in longer context scenarios. The default value is 8.",
        )
1078
1079
1080
1081
1082
1083
1084
1085
        parser.add_argument(
            "--num-continuous-decode-steps",
            type=int,
            default=ServerArgs.num_continuous_decode_steps,
            help="Run multiple continuous decoding steps to reduce scheduling overhead. "
            "This can potentially increase throughput but may also increase time-to-first-token latency. "
            "The default value is 1, meaning only run one decoding step at a time.",
        )
1086
1087
1088
1089
1090
        parser.add_argument(
            "--delete-ckpt-after-loading",
            action="store_true",
            help="Delete the model checkpoint after loading the model.",
        )
1091
1092
1093
1094
1095
        parser.add_argument(
            "--enable-memory-saver",
            action="store_true",
            help="Allow saving memory using release_memory_occupation and resume_memory_occupation",
        )
1096
1097
1098
1099
1100
        parser.add_argument(
            "--allow-auto-truncate",
            action="store_true",
            help="Allow automatically truncating requests that exceed the maximum input length instead of returning an error.",
        )
1101
1102
1103
1104
1105
        parser.add_argument(
            "--enable-custom-logit-processor",
            action="store_true",
            help="Enable users to pass custom logit processors to the server (disabled by default for security)",
        )
YAMY's avatar
YAMY committed
1106
1107
1108
        parser.add_argument(
            "--tool-call-parser",
            type=str,
1109
            choices=["qwen25", "mistral", "llama3", "deepseekv3"],
YAMY's avatar
YAMY committed
1110
1111
1112
            default=ServerArgs.tool_call_parser,
            help="Specify the parser for handling tool-call interactions. Options include: 'qwen25', 'mistral', and 'llama3'.",
        )
1113
1114
1115
1116
1117
        parser.add_argument(
            "--enable-hierarchical-cache",
            action="store_true",
            help="Enable hierarchical cache",
        )
1118
1119
1120
1121
1122
1123
        parser.add_argument(
            "--hicache-ratio",
            type=float,
            default=ServerArgs.hicache_ratio,
            help="The ratio of the size of host KV cache memory pool to the size of device pool.",
        )
Zhiqiang Xie's avatar
Zhiqiang Xie committed
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
        parser.add_argument(
            "--hicache-size",
            type=int,
            default=ServerArgs.hicache_size,
            help="The size of host KV cache memory pool in gigabytes, which will override the hicache_ratio if set.",
        )
        parser.add_argument(
            "--hicache-write-policy",
            type=str,
            choices=["write_back", "write_through", "write_through_selective"],
            default=ServerArgs.hicache_write_policy,
            help="The write policy of hierarchical cache.",
        )
1137
1138
1139
1140
1141
        parser.add_argument(
            "--enable-deepep-moe",
            action="store_true",
            help="Enabling DeepEP MoE implementation for EP MoE.",
        )
1142
1143
1144
1145
1146
1147
        parser.add_argument(
            "--moe-dense-tp-size",
            type=int,
            default=ServerArgs.moe_dense_tp_size,
            help="TP size for MoE dense MLP layers. This flag is useful when, with large TP size, there are errors caused by weights in MLP layers having dimension smaller than the min dimension GEMM supports.",
        )
1148
1149
1150
1151
        parser.add_argument(
            "--deepep-mode",
            type=str,
            choices=["normal", "low_latency", "auto"],
1152
            default="auto",
1153
1154
            help="Select the mode when enable DeepEP MoE, could be `normal`, `low_latency` or `auto`. Default is `auto`, which means `low_latency` for decode batch and `normal` for prefill batch.",
        )
1155

1156
1157
1158
        parser.add_argument(
            "--n-share-experts-fusion",
            type=int,
1159
            default=0,
1160
1161
            help="The number of shared_experts need to be replicated to fuse with normal experts in deepseek v3/r1, "
            "set it to tp_size can get best optimized performace.",
1162
        )
1163
1164
1165
1166
1167
        parser.add_argument(
            "--disable-chunked-prefix-cache",
            action="store_true",
            help="Disable chunked prefix cache feature for deepseek, which should save overhead for short sequences.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1168
1169
1170
1171
1172
        parser.add_argument(
            "--disable-fast-image-processor",
            action="store_true",
            help="Adopt base image processor instead of fast image processor.",
        )
1173

1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
        # Server warmups
        parser.add_argument(
            "--warmups",
            type=str,
            required=False,
            help="Specify custom warmup functions (csv) to run before server starts eg. --warmups=warmup_name1,warmup_name2 "
            "will run the functions `warmup_name1` and `warmup_name2` specified in warmup.py before the server starts listening for requests",
        )

        # Debug tensor dumps
        parser.add_argument(
            "--debug-tensor-dump-output-folder",
            type=str,
            default=ServerArgs.debug_tensor_dump_output_folder,
            help="The output folder for dumping tensors.",
        )
        parser.add_argument(
            "--debug-tensor-dump-input-file",
            type=str,
            default=ServerArgs.debug_tensor_dump_input_file,
            help="The input filename for dumping tensors",
        )
        parser.add_argument(
            "--debug-tensor-dump-inject",
            type=str,
            default=ServerArgs.debug_tensor_dump_inject,
            help="Inject the outputs from jax as the input of every layer.",
        )

Byron Hsu's avatar
Byron Hsu committed
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
        # Disaggregation
        parser.add_argument(
            "--disaggregation-mode",
            type=str,
            default="null",
            choices=["null", "prefill", "decode"],
            help='Only used for PD disaggregation. "prefill" for prefill-only server, and "decode" for decode-only server. If not specified, it is not PD disaggregated',
        )
        parser.add_argument(
            "--disaggregation-bootstrap-port",
            type=int,
            default=ServerArgs.disaggregation_bootstrap_port,
            help="Bootstrap server port on the prefill server. Default is 8998.",
        )
1217
1218
1219
1220
1221
1222
        parser.add_argument(
            "--disaggregation-transfer-backend",
            type=str,
            default=ServerArgs.disaggregation_transfer_backend,
            help="The backend for disaggregation transfer. Default is mooncake.",
        )
1223
1224
1225
1226
1227
1228
        parser.add_argument(
            "--disaggregation-ib-device",
            type=str,
            default=ServerArgs.disaggregation_ib_device,
            help="The ib device for disaggregation transfer. Default is None, it will be detected automatically if using the mooncake backend.",
        )
Byron Hsu's avatar
Byron Hsu committed
1229

Lianmin Zheng's avatar
Lianmin Zheng committed
1230
1231
    @classmethod
    def from_cli_args(cls, args: argparse.Namespace):
1232
1233
        args.tp_size = args.tensor_parallel_size
        args.dp_size = args.data_parallel_size
xiaobochen's avatar
xiaobochen committed
1234
        args.ep_size = args.expert_parallel_size
Lianmin Zheng's avatar
Lianmin Zheng committed
1235
1236
1237
1238
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        return cls(**{attr: getattr(args, attr) for attr in attrs})

    def url(self):
1239
        if is_valid_ipv6_address(self.host):
1240
1241
1242
            return f"http://[{self.host}]:{self.port}"
        else:
            return f"http://{self.host}:{self.port}"
Lianmin Zheng's avatar
Lianmin Zheng committed
1243

1244
1245
1246
1247
1248
    def check_server_args(self):
        assert (
            self.tp_size % self.nnodes == 0
        ), "tp_size must be divisible by number of nodes"
        assert not (
1249
1250
            self.dp_size > 1 and self.nnodes != 1 and not self.enable_dp_attention
        ), "multi-node data parallel is not supported unless dp attention!"
1251
1252
1253
1254
1255
1256
        assert (
            self.max_loras_per_batch > 0
            # FIXME
            and (self.lora_paths is None or self.disable_cuda_graph)
            and (self.lora_paths is None or self.disable_radix_cache)
        ), "compatibility of lora and cuda graph and radix attention is in progress"
1257
        assert self.base_gpu_id >= 0, "base_gpu_id must be non-negative"
1258
        assert self.gpu_id_step >= 1, "gpu_id_step must be positive"
1259

1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
        if isinstance(self.lora_paths, list):
            lora_paths = self.lora_paths
            self.lora_paths = {}
            for lora_path in lora_paths:
                if "=" in lora_path:
                    name, path = lora_path.split("=", 1)
                    self.lora_paths[name] = path
                else:
                    self.lora_paths[lora_path] = lora_path

Lianmin Zheng's avatar
Lianmin Zheng committed
1270

Lianmin Zheng's avatar
Lianmin Zheng committed
1271
def prepare_server_args(argv: List[str]) -> ServerArgs:
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
    """
    Prepare the server arguments from the command line arguments.

    Args:
        args: The command line arguments. Typically, it should be `sys.argv[1:]`
            to ensure compatibility with `parse_args` when no arguments are passed.

    Returns:
        The server arguments.
    """
    parser = argparse.ArgumentParser()
    ServerArgs.add_cli_args(parser)
Lianmin Zheng's avatar
Lianmin Zheng committed
1284
    raw_args = parser.parse_args(argv)
1285
1286
1287
1288
    server_args = ServerArgs.from_cli_args(raw_args)
    return server_args


1289
1290
1291
ZMQ_TCP_PORT_DELTA = 233


Lianmin Zheng's avatar
Lianmin Zheng committed
1292
1293
@dataclasses.dataclass
class PortArgs:
1294
1295
1296
1297
1298
1299
    # The ipc filename for tokenizer to receive inputs from detokenizer (zmq)
    tokenizer_ipc_name: str
    # The ipc filename for scheduler (rank 0) to receive inputs from tokenizer (zmq)
    scheduler_input_ipc_name: str
    # The ipc filename for detokenizer to receive inputs from scheduler (zmq)
    detokenizer_ipc_name: str
1300

1301
1302
    # The port for nccl initialization (torch.dist)
    nccl_port: int
1303

1304
1305
1306
    # The ipc filename for rpc call between Engine and Scheduler
    rpc_ipc_name: str

1307
    @staticmethod
1308
    def init_new(server_args, dp_rank: Optional[int] = None) -> "PortArgs":
1309
        port = server_args.port + random.randint(100, 1000)
1310
1311
1312
        while True:
            if is_port_available(port):
                break
TianYu GUO's avatar
TianYu GUO committed
1313
1314
1315
1316
            if port < 60000:
                port += 42
            else:
                port -= 43
1317

1318
1319
1320
1321
1322
1323
1324
        if not server_args.enable_dp_attention:
            # Normal case, use IPC within a single node
            return PortArgs(
                tokenizer_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
                scheduler_input_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
                detokenizer_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
                nccl_port=port,
1325
                rpc_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
1326
1327
1328
1329
1330
            )
        else:
            # DP attention. Use TCP + port to handle both single-node and multi-node.
            if server_args.nnodes == 1 and server_args.dist_init_addr is None:
                dist_init_addr = ("127.0.0.1", server_args.port + ZMQ_TCP_PORT_DELTA)
Vincent's avatar
Vincent committed
1331
1332
1333
            elif server_args.dist_init_addr.startswith("["):  # ipv6 address
                port_num, host = configure_ipv6(server_args.dist_init_addr)
                dist_init_addr = (host, str(port_num))
1334
1335
            else:
                dist_init_addr = server_args.dist_init_addr.split(":")
Vincent's avatar
Vincent committed
1336

1337
1338
1339
1340
1341
1342
1343
1344
            assert (
                len(dist_init_addr) == 2
            ), "please provide --dist-init-addr as host:port of head node"

            dist_init_host, dist_init_port = dist_init_addr
            port_base = int(dist_init_port) + 1
            if dp_rank is None:
                scheduler_input_port = (
1345
                    port_base + 3
1346
                )  # TokenizerManager to DataParallelController
1347
            else:
1348
                scheduler_input_port = port_base + 3 + 1 + dp_rank
1349
1350
1351
1352
1353
1354

            return PortArgs(
                tokenizer_ipc_name=f"tcp://{dist_init_host}:{port_base}",
                scheduler_input_ipc_name=f"tcp://{dist_init_host}:{scheduler_input_port}",
                detokenizer_ipc_name=f"tcp://{dist_init_host}:{port_base + 1}",
                nccl_port=port,
1355
                rpc_ipc_name=f"tcp://{dist_init_host}:{port_base + 2}",
1356
            )
1357

1358
1359
1360
1361
1362
1363
1364
1365
1366
1367

class LoRAPathAction(argparse.Action):
    def __call__(self, parser, namespace, values, option_string=None):
        setattr(namespace, self.dest, {})
        for lora_path in values:
            if "=" in lora_path:
                name, path = lora_path.split("=", 1)
                getattr(namespace, self.dest)[name] = path
            else:
                getattr(namespace, self.dest)[lora_path] = lora_path
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377


class DeprecatedAction(argparse.Action):
    def __init__(self, option_strings, dest, nargs=0, **kwargs):
        super(DeprecatedAction, self).__init__(
            option_strings, dest, nargs=nargs, **kwargs
        )

    def __call__(self, parser, namespace, values, option_string=None):
        raise ValueError(self.help)
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407


def auto_choose_speculative_params(self: ServerArgs):
    """
    Automatically choose the parameters for speculative decoding.

    You can tune them on your own models and prompts with scripts/playground/bench_speculative.py
    """
    if self.decrypted_config_file:
        config_path = self.decrypted_config_file
    else:
        config_path = os.path.join(self.model_path, "config.json")
    if not os.path.exists(config_path):
        raise ValueError(f"{config_path} is not found.")

    config = json.load(open(config_path))

    arch = config.get("architectures", ["Unknown"])[0]

    if arch in ["LlamaForCausalLM"]:
        # The default value for llama
        return (5, 4, 8)
    elif arch in ["DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"]:
        # The default value for deepseek
        return (5, 4, 8)
    elif arch in ["Grok1ForCausalLM", "Grok1VForCausalLM"]:
        return (5, 4, 8)
    else:
        # The default value for all other models
        return (5, 4, 8)