server_args.py 106 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 sys
23
import tempfile
24
from typing import List, Literal, Optional, Union
Lianmin Zheng's avatar
Lianmin Zheng committed
25

26
from sglang.srt.function_call.function_call_parser import FunctionCallParser
27
from sglang.srt.hf_transformers_utils import check_gguf_file, get_config
28
from sglang.srt.lora.lora_registry import LoRARef
29
from sglang.srt.parser.reasoning_parser import ReasoningParser
30
from sglang.srt.utils import (
31
32
    LORA_TARGET_ALL_MODULES,
    SUPPORTED_LORA_TARGET_MODULES,
Vincent's avatar
Vincent committed
33
    configure_ipv6,
34
    get_device,
Lianmin Zheng's avatar
Lianmin Zheng committed
35
    get_device_memory_capacity,
36
    is_cuda,
37
    is_flashinfer_available,
HAI's avatar
HAI committed
38
    is_hip,
39
    is_port_available,
40
    is_remote_url,
41
42
    is_sm90_supported,
    is_sm100_supported,
43
    is_triton_kernels_available,
44
    is_valid_ipv6_address,
bjmsong's avatar
bjmsong committed
45
    nullable_str,
46
)
47

48
49
logger = logging.getLogger(__name__)

Lianmin Zheng's avatar
Lianmin Zheng committed
50

51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
# Define constants
LOAD_FORMAT_CHOICES = [
    "auto",
    "pt",
    "safetensors",
    "npcache",
    "dummy",
    "sharded_state",
    "gguf",
    "bitsandbytes",
    "layered",
    "remote",
]

QUANTIZATION_CHOICES = [
    "awq",
    "fp8",
    "gptq",
    "marlin",
    "gptq_marlin",
    "awq_marlin",
    "bitsandbytes",
    "gguf",
    "modelopt",
    "modelopt_fp4",
    "petit_nvfp4",
    "w8a8_int8",
    "w8a8_fp8",
    "moe_wna16",
    "qoq",
    "w4afp8",
    "mxfp4",
]

ATTENTION_BACKEND_CHOICES = [
    # Common
    "triton",
    "torch_native",
    # NVIDIA specific
    "cutlass_mla",
    "fa3",
    "flashinfer",
    "flashmla",
    "trtllm_mla",
    "trtllm_mha",
    "dual_chunk_flash_attn",
    # AMD specific
    "aiter",
    "wave",
    # Other platforms
    "intel_amx",
    "ascend",
]

DISAGG_TRANSFER_BACKEND_CHOICES = ["mooncake", "nixl", "ascend", "fake"]

107
108
GRAMMAR_BACKEND_CHOICES = ["xgrammar", "outlines", "llguidance", "none"]

109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126

# Allow external code to add more choices
def add_load_format_choices(choices):
    LOAD_FORMAT_CHOICES.extend(choices)


def add_quantization_method_choices(choices):
    QUANTIZATION_CHOICES.extend(choices)


def add_attention_backend_choices(choices):
    ATTENTION_BACKEND_CHOICES.extend(choices)


def add_disagg_transfer_backend_choices(choices):
    DISAGG_TRANSFER_BACKEND_CHOICES.extend(choices)


127
128
129
130
def add_grammar_backend_choices(choices):
    GRAMMAR_BACKEND_CHOICES.extend(choices)


Lianmin Zheng's avatar
Lianmin Zheng committed
131
132
@dataclasses.dataclass
class ServerArgs:
Lianmin Zheng's avatar
Lianmin Zheng committed
133
    # Model and tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
134
135
136
    model_path: str
    tokenizer_path: Optional[str] = None
    tokenizer_mode: str = "auto"
137
    tokenizer_worker_num: int = 1
138
    skip_tokenizer_init: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
139
    load_format: str = "auto"
140
    model_loader_extra_config: str = "{}"
141
    trust_remote_code: bool = False
142
    context_length: Optional[int] = None
143
    is_embedding: bool = False
144
    enable_multimodal: Optional[bool] = None
145
    revision: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
146
    model_impl: str = "auto"
Lianmin Zheng's avatar
Lianmin Zheng committed
147

Lianmin Zheng's avatar
Lianmin Zheng committed
148
    # HTTP server
Lianmin Zheng's avatar
Lianmin Zheng committed
149
150
    host: str = "127.0.0.1"
    port: int = 30000
Lianmin Zheng's avatar
Lianmin Zheng committed
151
152
    skip_server_warmup: bool = False
    warmups: Optional[str] = None
153
    nccl_port: Optional[int] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
154

Lianmin Zheng's avatar
Lianmin Zheng committed
155
156
157
158
159
160
    # Quantization and data type
    dtype: str = "auto"
    quantization: Optional[str] = None
    quantization_param_path: Optional[str] = None
    kv_cache_dtype: str = "auto"

Lianmin Zheng's avatar
Lianmin Zheng committed
161
    # Memory and scheduling
Lianmin Zheng's avatar
Lianmin Zheng committed
162
    mem_fraction_static: Optional[float] = None
163
    max_running_requests: Optional[int] = None
164
    max_queued_requests: Optional[int] = sys.maxsize
165
    max_total_tokens: Optional[int] = None
166
    chunked_prefill_size: Optional[int] = None
167
    max_prefill_tokens: int = 16384
168
    schedule_policy: str = "fcfs"
169
    schedule_conservativeness: float = 1.0
170
    page_size: Optional[int] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
171
172
173
    hybrid_kvcache_ratio: Optional[float] = None
    swa_full_tokens_ratio: float = 0.8
    disable_hybrid_swa_memory: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
174

Lianmin Zheng's avatar
Lianmin Zheng committed
175
176
    # Runtime options
    device: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
177
    tp_size: int = 1
178
179
    pp_size: int = 1
    max_micro_batch_size: Optional[int] = None
180
    stream_interval: int = 1
181
    stream_output: bool = False
182
    random_seed: Optional[int] = None
183
    constrained_json_whitespace_pattern: Optional[str] = None
184
    watchdog_timeout: float = 300
185
    dist_timeout: Optional[int] = None  # timeout for torch.distributed
186
    download_dir: Optional[str] = None
187
    base_gpu_id: int = 0
188
    gpu_id_step: int = 1
189
    sleep_on_idle: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
190
191
192

    # Logging
    log_level: str = "info"
193
    log_level_http: Optional[str] = None
194
    log_requests: bool = False
195
    log_requests_level: int = 2
196
    crash_dump_folder: Optional[str] = None
Liangsheng Yin's avatar
Liangsheng Yin committed
197
    show_time_cost: bool = False
198
    enable_metrics: bool = False
199
    enable_metrics_for_all_schedulers: bool = False
200
201
    bucket_time_to_first_token: Optional[List[float]] = None
    bucket_inter_token_latency: Optional[List[float]] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
202
    bucket_e2e_request_latency: Optional[List[float]] = None
203
    collect_tokens_histogram: bool = False
204
205
    prompt_tokens_buckets: Optional[List[str]] = None
    generation_tokens_buckets: Optional[List[str]] = None
206
    decode_log_interval: int = 40
207
    enable_request_time_stats_logging: bool = False
208
    kv_events_config: Optional[str] = None
209
    gc_warning_threshold_secs: float = 0.0
Liangsheng Yin's avatar
Liangsheng Yin committed
210

211
    # API related
212
    api_key: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
213
    served_model_name: Optional[str] = None
214
    weight_version: str = "default"
Lianmin Zheng's avatar
Lianmin Zheng committed
215
216
    chat_template: Optional[str] = None
    completion_template: Optional[str] = None
217
    file_storage_path: str = "sglang_storage"
218
    enable_cache_report: bool = False
Xihuai Wang's avatar
Xihuai Wang committed
219
    reasoning_parser: Optional[str] = None
220
    tool_call_parser: Optional[str] = None
221
    tool_server: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
222

223
224
225
    # Data parallelism
    dp_size: int = 1
    load_balance_method: str = "round_robin"
226

227
    # Multi-node distributed serving
228
    dist_init_addr: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
229
    nnodes: int = 1
230
    node_rank: int = 0
Lianmin Zheng's avatar
Lianmin Zheng committed
231
232
233

    # Model override args in JSON
    json_model_override_args: str = "{}"
234
    preferred_sampling_params: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
235

236
    # LoRA
237
    enable_lora: Optional[bool] = None
238
    max_lora_rank: Optional[int] = None
239
    lora_target_modules: Optional[Union[set[str], List[str]]] = None
240
241
242
    lora_paths: Optional[
        Union[dict[str, str], List[dict[str, str]], List[str], List[LoRARef]]
    ] = None
243
    max_loaded_loras: Optional[int] = None
244
    max_loras_per_batch: int = 8
245
    lora_backend: str = "triton"
246
247

    # Kernel backend
248
    attention_backend: Optional[str] = None
249
250
    decode_attention_backend: Optional[str] = None
    prefill_attention_backend: Optional[str] = None
251
    sampling_backend: Optional[str] = None
252
    grammar_backend: Optional[str] = None
253
    mm_attention_backend: Optional[str] = None
254

255
256
    # Speculative decoding
    speculative_algorithm: Optional[str] = None
257
    speculative_draft_model_path: Optional[str] = None
258
    speculative_draft_model_revision: Optional[str] = None
259
260
261
    speculative_num_steps: Optional[int] = None
    speculative_eagle_topk: Optional[int] = None
    speculative_num_draft_tokens: Optional[int] = None
262
263
    speculative_accept_threshold_single: float = 1.0
    speculative_accept_threshold_acc: float = 1.0
264
    speculative_token_map: Optional[str] = None
265

266
267
    # Expert parallelism
    ep_size: int = 1
268
269
270
271
272
273
274
275
276
    moe_a2a_backend: Literal["none", "deepep"] = "none"
    moe_runner_backend: Literal[
        "auto",
        "triton",
        "triton_kernel",
        "flashinfer_trtllm",
        "flashinfer_cutlass",
        "flashinfer_mxfp4",
    ] = "auto"
277
    flashinfer_mxfp4_moe_precision: Literal["default", "bf16"] = "default"
278
    enable_flashinfer_allreduce_fusion: bool = False
279
    deepep_mode: Literal["auto", "normal", "low_latency"] = "auto"
280
281
282
283
284
285
286
    ep_num_redundant_experts: int = 0
    ep_dispatch_algorithm: Optional[Literal["static", "dynamic", "fake"]] = None
    init_expert_location: str = "trivial"
    enable_eplb: bool = False
    eplb_algorithm: str = "auto"
    eplb_rebalance_num_iterations: int = 1000
    eplb_rebalance_layers_per_chunk: Optional[int] = None
287
    eplb_min_rebalancing_utilization_threshold: float = 1.0
288
289
290
291
292
293
294
295
    expert_distribution_recorder_mode: Optional[
        Literal["stat", "stat_approx", "per_pass", "per_token"]
    ] = None
    expert_distribution_recorder_buffer_size: Optional[int] = None
    enable_expert_distribution_metrics: bool = False
    deepep_config: Optional[str] = None
    moe_dense_tp_size: Optional[int] = None

Lianmin Zheng's avatar
Lianmin Zheng committed
296
297
298
299
    # Hierarchical cache
    enable_hierarchical_cache: bool = False
    hicache_ratio: float = 2.0
    hicache_size: int = 0
300
    hicache_write_policy: str = "write_through"
301
302
    hicache_io_backend: str = "kernel"
    hicache_mem_layout: str = "layer_first"
Lianmin Zheng's avatar
Lianmin Zheng committed
303
    hicache_storage_backend: Optional[str] = None
pansicheng's avatar
pansicheng committed
304
    hicache_storage_prefetch_policy: str = "best_effort"
305
    hicache_storage_backend_extra_config: Optional[str] = None
306
307
    # LMCache
    enable_lmcache: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
308

309
310
    # Double Sparsity
    enable_double_sparsity: bool = False
Vincent's avatar
Vincent committed
311
    ds_channel_config_path: Optional[str] = None
312
313
314
315
316
    ds_heavy_channel_num: int = 32
    ds_heavy_token_num: int = 256
    ds_heavy_channel_type: str = "qk"
    ds_sparse_decode_threshold: int = 4096

fzyzcjy's avatar
fzyzcjy committed
317
318
319
320
321
322
323
    # Offloading
    cpu_offload_gb: int = 0
    offload_group_size: int = -1
    offload_num_in_group: int = 1
    offload_prefetch_step: int = 1
    offload_mode: str = "cpu"

324
    # Optimization/debug options
Lianmin Zheng's avatar
Lianmin Zheng committed
325
    disable_radix_cache: bool = False
326
327
    cuda_graph_max_bs: Optional[int] = None
    cuda_graph_bs: Optional[List[int]] = None
328
    disable_cuda_graph: bool = False
329
    disable_cuda_graph_padding: bool = False
330
    enable_profile_cuda_graph: bool = False
331
    enable_cudagraph_gc: bool = False
332
    enable_nccl_nvls: bool = False
333
    enable_symm_mem: bool = False
334
    disable_flashinfer_cutlass_moe_fp4_allgather: bool = False
335
    enable_tokenizer_batch_encode: bool = False
336
    disable_outlines_disk_cache: bool = False
337
    disable_custom_all_reduce: bool = False
338
    enable_mscclpp: bool = False
339
    disable_overlap_schedule: bool = False
340
    enable_mixed_chunk: bool = False
Ke Bao's avatar
Ke Bao committed
341
    enable_dp_attention: bool = False
342
    enable_dp_lm_head: bool = False
343
    enable_two_batch_overlap: bool = False
344
    tbo_token_distribution_threshold: float = 0.48
345
    enable_torch_compile: bool = False
346
    torch_compile_max_bs: int = 32
347
    torchao_config: str = ""
348
    enable_nan_detection: bool = False
349
    enable_p2p_check: bool = False
350
    triton_attention_reduce_in_fp32: bool = False
351
    triton_attention_num_kv_splits: int = 8
352
    num_continuous_decode_steps: int = 1
353
    delete_ckpt_after_loading: bool = False
354
    enable_memory_saver: bool = False
355
    allow_auto_truncate: bool = False
356
    enable_custom_logit_processor: bool = False
357
    flashinfer_mla_disable_ragged: bool = False
358
    disable_shared_experts_fusion: bool = False
359
    disable_chunked_prefix_cache: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
360
    disable_fast_image_processor: bool = False
361
    enable_return_hidden_states: bool = False
362
    scheduler_recv_interval: int = 1
363
    numa_node: Optional[List[int]] = None
364
365
366
367
368

    # 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
369
    debug_tensor_dump_prefill_only: bool = False
370

Lianmin Zheng's avatar
Lianmin Zheng committed
371
    # PD disaggregation: can be "null" (not disaggregated), "prefill" (prefill-only), or "decode" (decode-only)
Byron Hsu's avatar
Byron Hsu committed
372
    disaggregation_mode: str = "null"
373
    disaggregation_transfer_backend: str = "mooncake"
374
    disaggregation_bootstrap_port: int = 8998
Byron Hsu's avatar
Byron Hsu committed
375
376
377
    disaggregation_decode_tp: Optional[int] = None
    disaggregation_decode_dp: Optional[int] = None
    disaggregation_prefill_pp: Optional[int] = 1
378
    disaggregation_ib_device: Optional[str] = None
379
    num_reserved_decode_tokens: int = 512  # used for decode kv cache offload in PD
Byron Hsu's avatar
Byron Hsu committed
380

381
382
    # For model weight update
    custom_weight_loader: Optional[List[str]] = None
383
    weight_loader_disable_mmap: bool = False
384

385
386
387
388
    # For PD-Multiplexing
    enable_pdmux: bool = False
    sm_group_num: int = 3

389
390
391
    # Deprecated arguments
    enable_ep_moe: bool = False
    enable_deepep_moe: bool = False
392
393
394
    enable_flashinfer_cutlass_moe: bool = False
    enable_flashinfer_trtllm_moe: bool = False
    enable_triton_kernel_moe: bool = False
395
    enable_flashinfer_mxfp4_moe: bool = False
396

Lianmin Zheng's avatar
Lianmin Zheng committed
397
    def __post_init__(self):
398
399
400
401
402
403
404
405
406
407
408
        # Check deprecated arguments
        if self.enable_ep_moe:
            self.ep_size = self.tp_size
            print_deprecated_warning(
                "NOTE: --enable-ep-moe is deprecated. Please set `--ep-size` to the same value as `--tp-size` instead."
            )
        if self.enable_deepep_moe:
            self.moe_a2a_backend = "deepep"
            print_deprecated_warning(
                "NOTE: --enable-deepep-moe is deprecated. Please set `--moe-a2a-backend` to 'deepep' instead."
            )
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
        if self.enable_triton_kernel_moe:
            self.moe_runner_backend = "triton_kernel"
            print_deprecated_warning(
                "NOTE: --enable-triton-kernel-moe is deprecated. Please set `--moe-runner-backend` to 'triton_kernel' instead."
            )
        if self.enable_flashinfer_cutlass_moe:
            self.moe_runner_backend = "flashinfer_cutlass"
            print_deprecated_warning(
                "NOTE: --enable-flashinfer-cutlass-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_cutlass' instead."
            )
        if self.enable_flashinfer_trtllm_moe:
            self.moe_runner_backend = "flashinfer_trtllm"
            print_deprecated_warning(
                "NOTE: --enable-flashinfer-trtllm-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_trtllm' instead."
            )
424
425
426
427
428
        if self.enable_flashinfer_mxfp4_moe:
            self.moe_runner_backend = "flashinfer_mxfp4"
            print_deprecated_warning(
                "NOTE: --enable-flashinfer-mxfp4-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_mxfp4' instead."
            )
429

430
        # Set missing default values
Lianmin Zheng's avatar
Lianmin Zheng committed
431
432
        if self.tokenizer_path is None:
            self.tokenizer_path = self.model_path
433
434
        if self.served_model_name is None:
            self.served_model_name = self.model_path
435
436
        if self.device is None:
            self.device = get_device()
437
438
439
        if self.random_seed is None:
            self.random_seed = random.randint(0, 1 << 30)

Lianmin Zheng's avatar
Lianmin Zheng committed
440
        gpu_mem = get_device_memory_capacity(self.device)
441

442
        # Set mem fraction static
Lianmin Zheng's avatar
Lianmin Zheng committed
443
        if self.mem_fraction_static is None:
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
            if gpu_mem is not None:
                # GPU memory capacity = model weights + KV cache pool + activations + cuda graph buffers
                # mem_fraction_static = (model weights + KV cache pool) / GPU memory capacity.

                # We want mem_fraction_static to be as large as possible but still has enough room
                # for activations and cuda graph buffers. We use the following heuristic to
                # compute the needed size for activations and cuda graph buffers:
                # - The size of the activation depends on the chunked_prefill_size and model size.
                # - The size of cuda graph buffers depends on the cuda graph capture range and model size.
                # For GPUs with more memory, we use a larger chunked_prefill_size and
                # capture more cuda graphs, so they need to reserve more memory.
                parallel_size = self.tp_size * self.pp_size

                if gpu_mem < 20 * 1024:
                    # T4, 4080. (chunked_prefill_size 2k, cuda_graph_max_bs 8)
                    reserved_mem = (2.8 + parallel_size / 10) * 1024
                elif gpu_mem < 35 * 1024:
                    # A10, L40, 4090, 5090. (chunked_prefill_size 2k, cuda_graph_max_bs 8)
                    reserved_mem = (2.8 + parallel_size / 10) * 1024
                elif gpu_mem < 90 * 1024:
                    # H100, A100. (chunked_prefill_size 8k, cuda_graph_max_bs 160)
                    reserved_mem = (9.5 + parallel_size / 2) * 1024
                elif gpu_mem < 100 * 1024:
                    # H20. (chunked_prefill_size 8k, cuda_graph_max_bs 256)
                    reserved_mem = (12 + parallel_size / 2) * 1024
                elif gpu_mem < 160 * 1024:
                    # H200. (chunked_prefill_size 8k, cuda_graph_max_bs 256)
                    reserved_mem = (12 + parallel_size / 2) * 1024
472
                else:
473
474
475
                    # B200, MI300. (chunked_prefill_size 16k, cuda_graph_max_bs 512)
                    reserved_mem = 32 * 1024

476
                if self.speculative_algorithm is not None:
477
478
479
480
481
482
                    # draft model and larger cuda graph buffers
                    reserved_mem += 2 * 1024
                if self.enable_dp_attention:
                    reserved_mem += 4 * 1024

                self.mem_fraction_static = round((gpu_mem - reserved_mem) / gpu_mem, 3)
483
            else:
484
                self.mem_fraction_static = 0.88
485

486
            # Lazy init to avoid circular import
Lianmin Zheng's avatar
Lianmin Zheng committed
487
            # Multimodal models need more memory for the image processor
488
489
490
            from sglang.srt.configs.model_config import ModelConfig

            model_config = ModelConfig.from_server_args(self)
Lianmin Zheng's avatar
Lianmin Zheng committed
491
492
            if model_config.is_multimodal:
                self.adjust_mem_fraction_for_vlm(model_config)
493

494
495
        # Set chunked prefill size, which depends on the gpu memory capacity
        if self.chunked_prefill_size is None:
496
497
498
499
500
501
502
            if gpu_mem is not None:
                if gpu_mem < 35 * 1024:  # A10, L40, 4090
                    self.chunked_prefill_size = 2048
                elif gpu_mem < 160 * 1024:  # H100, H200, A100, H20
                    self.chunked_prefill_size = 8192
                else:  # B200, MI300
                    self.chunked_prefill_size = 16384
503
            else:
504
                self.chunked_prefill_size = 4096
Lianmin Zheng's avatar
Lianmin Zheng committed
505

506
507
508
509
510
511
512
513
514
        # Set cuda graph max batch size
        if self.cuda_graph_max_bs is None:
            # 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.
            if gpu_mem is not None and gpu_mem < 35 * 1024:
                if self.tp_size < 4:
                    self.cuda_graph_max_bs = 8
                else:
                    self.cuda_graph_max_bs = 80

515
        # Set kernel backends for hpu device
516
517
518
519
        if self.device == "hpu":
            self.attention_backend = "torch_native"
            self.sampling_backend = "pytorch"

520
521
522
        # Model-specific adjustments
        self.model_specific_adjustments()

Lianmin Zheng's avatar
Lianmin Zheng committed
523
        # Set kernel backends
524
525
526
527
528
        if self.device == "cpu":
            if self.attention_backend is None:
                self.attention_backend = "intel_amx"
            self.sampling_backend = "pytorch"

529
        if self.sampling_backend is None:
530
531
532
533
534
            self.sampling_backend = (
                "flashinfer" if is_flashinfer_available() else "pytorch"
            )

        if self.attention_backend == "torch_native":
535
            logger.warning(
536
537
538
                "Cuda graph is disabled because of using torch native attention backend"
            )
            self.disable_cuda_graph = True
539

540
541
542
543
544
545
        if self.attention_backend == "ascend":
            logger.warning(
                "At this moment Ascend attention backend only supports a page_size of 128, change page_size to 128."
            )
            self.page_size = 128

546
547
548
549
        if (
            self.attention_backend == "flashmla"
            or self.decode_attention_backend == "flashmla"
        ):
Lianmin Zheng's avatar
Lianmin Zheng committed
550
551
552
553
554
            logger.warning(
                "FlashMLA only supports a page_size of 64, change page_size to 64."
            )
            self.page_size = 64

555
556
557
558
        if (
            self.attention_backend == "cutlass_mla"
            or self.decode_attention_backend == "cutlass_mla"
        ):
Lianmin Zheng's avatar
Lianmin Zheng committed
559
560
561
562
563
            logger.warning(
                "Cutlass MLA only supports a page_size of 128, change page_size to 128."
            )
            self.page_size = 128

Faraz's avatar
Faraz committed
564
565
566
567
        if (
            self.attention_backend == "trtllm_mla"
            or self.decode_attention_backend == "trtllm_mla"
        ):
568
569
570
571
572
573
574
575
576
577
            if not is_sm100_supported():
                raise ValueError(
                    "TRTLLM MLA backend is only supported on Blackwell GPUs (SM100). Please use a different backend."
                )

            if self.page_size not in [32, 64]:
                logger.warning(
                    f"TensorRT-LLM MLA only supports page_size of 32 or 64, changing page_size from {self.page_size} to 64."
                )
                self.page_size = 64
Faraz's avatar
Faraz committed
578
579
580
581
582

            if self.kv_cache_dtype not in ["fp8_e4m3", "auto"]:
                raise ValueError(
                    "TensorRT-LLM MLA backend only supports kv-cache-dtype of fp8_e4m3 or auto."
                )
583

584
585
586
587
588
        if (
            self.attention_backend == "trtllm_mha"
            or self.decode_attention_backend == "trtllm_mha"
            or self.prefill_attention_backend == "trtllm_mha"
        ):
589
590
591
592
593
594
595
596
597
598
599
            if not is_sm100_supported():
                raise ValueError(
                    "TRTLLM MHA backend is only supported on Blackwell GPUs (SM100). Please use a different backend."
                )

            if self.page_size not in [16, 32, 64]:
                logger.warning(
                    f"TensorRT-LLM MHA only supports page_size of 16, 32 or 64, changing page_size from {self.page_size} to 64."
                )
                self.page_size = 64

600
601
        if self.attention_backend == "dual_chunk_flash_attn":
            logger.warning(
602
                "Mixed chunk, radix cache, and cuda graphs are disabled because of using dual chunk flash attention backend"
603
604
605
606
607
            )
            self.enable_mixed_chunk = False
            self.disable_cuda_graph = True
            self.disable_radix_cache = True

608
609
610
611
612
613
614
615
        # Set page size
        if self.page_size is None:
            self.page_size = 1

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

616
617
618
        # Choose grammar backend
        if self.grammar_backend is None:
            self.grammar_backend = "xgrammar"
619

620
        # Data parallelism attention
Ke Bao's avatar
Ke Bao committed
621
        if self.enable_dp_attention:
622
            self.schedule_conservativeness = self.schedule_conservativeness * 0.3
Lianmin Zheng's avatar
Lianmin Zheng committed
623
624
625
626
627
            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
628
            logger.warning(
629
                f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
630
            )
631

632
633
634
        if self.enable_dp_lm_head:
            assert (
                self.enable_dp_attention
635
            ), "Please enable dp attention when setting enable_dp_lm_head. "
636

637
        # MoE kernel
638
        if self.moe_runner_backend == "flashinfer_cutlass":
639
640
641
            assert (
                self.quantization == "modelopt_fp4"
            ), "modelopt_fp4 quantization is required for Flashinfer MOE"
642
643
644
645
            assert self.ep_size in [
                1,
                self.tp_size,
            ], "The expert parallel size must be 1 or the same as the tensor parallel size"
646

647
        if self.moe_runner_backend == "flashinfer_trtllm":
648
649
650
651
652
653
            if not self.disable_shared_experts_fusion:
                self.disable_shared_experts_fusion = True
                logger.warning(
                    "FlashInfer TRTLLM MoE is enabled. --disable-shared-experts-fusion is automatically set."
                )

654
        # DeepEP MoE
655
        if self.moe_a2a_backend == "deepep":
656
657
658
            if self.deepep_mode == "normal":
                logger.warning("Cuda graph is disabled because deepep_mode=`normal`")
                self.disable_cuda_graph = True
659
            self.ep_size = self.tp_size
Lianmin Zheng's avatar
Lianmin Zheng committed
660
            logger.warning(
661
662
                f"DeepEP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
            )
663

664
665
        if self.enable_eplb and (self.expert_distribution_recorder_mode is None):
            self.expert_distribution_recorder_mode = "stat"
666
            logger.warning(
667
                "EPLB is enabled. The expert_distribution_recorder_mode is automatically set."
668
669
670
671
672
673
674
            )

        if (self.enable_eplb or (self.init_expert_location is not None)) and (
            self.ep_dispatch_algorithm is None
        ):
            self.ep_dispatch_algorithm = "static"

675
        if self.enable_eplb:
676
            assert self.ep_size > 1
677

678
679
680
681
682
        if self.enable_expert_distribution_metrics and (
            self.expert_distribution_recorder_mode is None
        ):
            self.expert_distribution_recorder_mode = "stat"

683
        if self.expert_distribution_recorder_buffer_size is None:
684
685
            if (x := self.eplb_rebalance_num_iterations) is not None:
                self.expert_distribution_recorder_buffer_size = x
686
687
688
            elif self.expert_distribution_recorder_mode is not None:
                self.expert_distribution_recorder_buffer_size = 1000

Lianmin Zheng's avatar
Lianmin Zheng committed
689
690
691
692
693
694
695
        # Pipeline parallelism
        if self.pp_size > 1:
            self.disable_overlap_schedule = True
            logger.warning(
                "Pipeline parallelism is incompatible with overlap schedule."
            )

Lianmin Zheng's avatar
Lianmin Zheng committed
696
        # Hicache
697
698
699
700
701
        if self.hicache_storage_backend == "mooncake":
            # to use mooncake storage backend, the following conditions must be met:
            self.hicache_io_backend = "kernel"
            self.hicache_mem_layout = "page_first"

702
        # Speculative Decoding
703
704
705
706
        if self.speculative_algorithm == "NEXTN":
            # NEXTN shares the same implementation of EAGLE
            self.speculative_algorithm = "EAGLE"

Lianmin Zheng's avatar
Lianmin Zheng committed
707
        if self.speculative_algorithm in ("EAGLE", "EAGLE3"):
708
            if self.max_running_requests is None:
709
                self.max_running_requests = 48
710
            self.disable_overlap_schedule = True
Lianmin Zheng's avatar
Lianmin Zheng committed
711
            logger.warning(
712
                "Overlap scheduler is disabled because of using "
713
                "eagle speculative decoding."
714
            )
715
716
717
718
719
720
            if self.enable_mixed_chunk:
                self.enable_mixed_chunk = False
                logger.warning(
                    "Mixed chunked prefill is disabled because of using "
                    "eagle speculative decoding."
                )
721

Lianmin Zheng's avatar
Lianmin Zheng committed
722
            model_arch = self.get_hf_config().architectures[0]
Yuxuan Zhang's avatar
Yuxuan Zhang committed
723
            if model_arch in ["DeepseekV3ForCausalLM", "Glm4MoeForCausalLM"]:
Hanming Lu's avatar
Hanming Lu committed
724
                # Auto set draft_model_path DeepSeek-V3/R1
725
726
727
728
729
730
                if self.speculative_draft_model_path is None:
                    self.speculative_draft_model_path = self.model_path
                else:
                    logger.warning(
                        "DeepSeek MTP does not require setting speculative_draft_model_path."
                    )
731

732
733
734
735
736
737
738
739
740
741
            # 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,
742
                ) = auto_choose_speculative_params(self)
743

744
745
746
747
748
749
750
751
752
753
            if (
                self.attention_backend == "trtllm_mha"
                or self.decode_attention_backend == "trtllm_mha"
                or self.prefill_attention_backend == "trtllm_mha"
            ):
                if self.speculative_eagle_topk > 1:
                    raise ValueError(
                        "trtllm_mha backend only supports topk = 1 for speculative decoding."
                    )

754
755
756
757
            if (
                self.speculative_eagle_topk == 1
                and self.speculative_num_draft_tokens != self.speculative_num_steps + 1
            ):
Lianmin Zheng's avatar
Lianmin Zheng committed
758
                logger.warning(
759
760
761
                    "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
762

763
764
765
766
767
768
769
770
771
            if (
                self.speculative_eagle_topk > 1
                and self.page_size > 1
                and self.attention_backend != "flashinfer"
            ):
                raise ValueError(
                    "speculative_eagle_topk > 1 with page_size > 1 is unstable and produces incorrect results for paged attention backends. This combination is only supported for the 'flashinfer' backend."
                )

772
            # The token generated from the verify step is counted.
773
            # If sepculative_num_steps >= speculative_num_draft_tokens, the additional tokens will definitely be discarded.
774
            # assert self.speculative_num_steps < self.speculative_num_draft_tokens
775

776
777
778
779
780
781
        # GGUF
        if (
            self.load_format == "auto" or self.load_format == "gguf"
        ) and check_gguf_file(self.model_path):
            self.quantization = self.load_format = "gguf"

782
        # Model loading
783
784
        if is_remote_url(self.model_path):
            self.load_format = "remote"
785
786
        if self.custom_weight_loader is None:
            self.custom_weight_loader = []
787

Byron Hsu's avatar
Byron Hsu committed
788
        # PD disaggregation
Byron Hsu's avatar
Byron Hsu committed
789
790
791
792
793
794
795
796
        if self.disaggregation_mode == "decode":
            assert (
                self.disaggregation_decode_tp is None
            ), "Cannot set --disaggregation-decode-tp for the decode engine."
            assert (
                self.disaggregation_decode_dp is None
            ), "Cannot set --disaggregation-decode-dp for the decode engine."

Byron Hsu's avatar
Byron Hsu committed
797
            self.disable_radix_cache = True
798
            logger.warning("KV cache is forced as chunk cache for decode server")
Byron Hsu's avatar
Byron Hsu committed
799
800
801
802
803
804
805
806
807
808
809
        elif self.disaggregation_mode == "prefill":
            if self.disaggregation_decode_tp is None:
                self.disaggregation_decode_tp = self.tp_size
            if self.disaggregation_decode_dp is None:
                self.disaggregation_decode_dp = self.dp_size

            self.disaggregation_prefill_pp = self.pp_size
            self.validate_disagg_tp_size(self.tp_size, self.disaggregation_decode_tp)

            self.disable_cuda_graph = True
            logger.warning("Cuda graph is disabled for prefill server")
Byron Hsu's avatar
Byron Hsu committed
810

811
        # Propagate env vars
812
813
814
        os.environ["SGLANG_ENABLE_TORCH_COMPILE"] = (
            "1" if self.enable_torch_compile else "0"
        )
815
816
817
818
        # Set env var before grammar backends init
        os.environ["SGLANG_DISABLE_OUTLINES_DISK_CACHE"] = (
            "1" if self.disable_outlines_disk_cache else "0"
        )
819

820
821
822
823
824
825
        if self.enable_hierarchical_cache and self.disable_radix_cache:
            raise ValueError(
                "The arguments enable-hierarchical-cache and disable-radix-cache are mutually exclusive "
                "and cannot be used at the same time. Please use only one of them."
            )

Lianmin Zheng's avatar
Lianmin Zheng committed
826
827
    @staticmethod
    def add_cli_args(parser: argparse.ArgumentParser):
Lianmin Zheng's avatar
Lianmin Zheng committed
828
        # Model and tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
829
830
        parser.add_argument(
            "--model-path",
831
            "--model",
Lianmin Zheng's avatar
Lianmin Zheng committed
832
833
834
835
836
837
838
839
840
841
            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.",
        )
842
843
844
845
846
847
        parser.add_argument(
            "--tokenizer-worker-num",
            type=int,
            default=ServerArgs.tokenizer_worker_num,
            help="The worker num of the tokenizer manager.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
848
849
850
851
852
853
854
855
856
        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.",
        )
857
858
859
        parser.add_argument(
            "--skip-tokenizer-init",
            action="store_true",
860
            help="If set, skip init tokenizer and pass input_ids in generate request.",
861
        )
862
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
863
864
865
            "--load-format",
            type=str,
            default=ServerArgs.load_format,
866
            choices=LOAD_FORMAT_CHOICES,
Lianmin Zheng's avatar
Lianmin Zheng committed
867
868
869
870
871
872
873
874
875
            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, '
876
            "which is mainly for profiling."
877
878
            '"gguf" will load the weights in the gguf format. '
            '"bitsandbytes" will load the weights using bitsandbytes '
879
880
881
882
            "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
883
        )
884
885
886
887
888
889
890
        parser.add_argument(
            "--model-loader-extra-config",
            type=str,
            help="Extra config for model loader. "
            "This will be passed to the model loader corresponding to the chosen load_format.",
            default=ServerArgs.model_loader_extra_config,
        )
891
892
893
894
895
        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
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
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
        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(
            "--is-embedding",
            action="store_true",
            help="Whether to use a CausalLM as an embedding model.",
        )
        parser.add_argument(
            "--enable-multimodal",
            default=ServerArgs.enable_multimodal,
            action="store_true",
            help="Enable the multimodal functionality for the served model. If the model being served is not multimodal, nothing will happen",
        )
        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.",
        )
        parser.add_argument(
            "--model-impl",
            type=str,
            default=ServerArgs.model_impl,
            help="Which implementation of the model to use.\n\n"
            '* "auto" will try to use the SGLang implementation if it exists '
            "and fall back to the Transformers implementation if no SGLang "
            "implementation is available.\n"
            '* "sglang" will use the SGLang model implementation.\n'
            '* "transformers" will use the Transformers model '
            "implementation.\n",
        )

        # HTTP server
        parser.add_argument(
            "--host",
            type=str,
            default=ServerArgs.host,
            help="The host of the HTTP server.",
        )
        parser.add_argument(
            "--port",
            type=int,
            default=ServerArgs.port,
            help="The port of the HTTP server.",
        )
        parser.add_argument(
            "--skip-server-warmup",
            action="store_true",
            help="If set, skip warmup.",
        )
        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",
        )
        parser.add_argument(
            "--nccl-port",
            type=int,
            default=ServerArgs.nccl_port,
            help="The port for NCCL distributed environment setup. Defaults to a random port.",
        )

        # Quantization and data type
Lianmin Zheng's avatar
Lianmin Zheng committed
967
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
968
            "--dtype",
Cody Yu's avatar
Cody Yu committed
969
            type=str,
Lianmin Zheng's avatar
Lianmin Zheng committed
970
            default=ServerArgs.dtype,
Ying Sheng's avatar
Ying Sheng committed
971
972
            choices=["auto", "half", "float16", "bfloat16", "float", "float32"],
            help="Data type for model weights and activations.\n\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
973
            '* "auto" will use FP16 precision for FP32 and FP16 models, and '
Ying Sheng's avatar
Ying Sheng committed
974
            "BF16 precision for BF16 models.\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
975
976
977
978
            '* "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
979
980
            '* "float32" for FP32 precision.',
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
981
982
983
984
        parser.add_argument(
            "--quantization",
            type=str,
            default=ServerArgs.quantization,
985
            choices=QUANTIZATION_CHOICES,
Lianmin Zheng's avatar
Lianmin Zheng committed
986
987
            help="The quantization method.",
        )
988
989
990
991
992
993
994
995
996
        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. ",
        )
997
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
998
            "--kv-cache-dtype",
999
            type=str,
Lianmin Zheng's avatar
Lianmin Zheng committed
1000
1001
1002
            default=ServerArgs.kv_cache_dtype,
            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+.',
1003
        )
1004

1005
        # Memory and scheduling
Lianmin Zheng's avatar
Lianmin Zheng committed
1006
1007
1008
1009
        parser.add_argument(
            "--mem-fraction-static",
            type=float,
            default=ServerArgs.mem_fraction_static,
1010
            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
1011
        )
1012
1013
1014
1015
1016
1017
        parser.add_argument(
            "--max-running-requests",
            type=int,
            default=ServerArgs.max_running_requests,
            help="The maximum number of running requests.",
        )
1018
1019
1020
1021
1022
1023
        parser.add_argument(
            "--max-queued-requests",
            type=int,
            default=ServerArgs.max_queued_requests,
            help="The maximum number of queued requests. This option is ignored when using disaggregation-mode.",
        )
1024
1025
1026
1027
        parser.add_argument(
            "--max-total-tokens",
            type=int,
            default=ServerArgs.max_total_tokens,
1028
1029
            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.",
1030
        )
1031
1032
1033
1034
        parser.add_argument(
            "--chunked-prefill-size",
            type=int,
            default=ServerArgs.chunked_prefill_size,
1035
            help="The maximum number of tokens in a chunk for the chunked prefill. Setting this to -1 means disabling chunked prefill.",
1036
1037
1038
1039
1040
1041
1042
        )
        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
1043
        parser.add_argument(
1044
            "--schedule-policy",
Lianmin Zheng's avatar
Lianmin Zheng committed
1045
            type=str,
1046
            default=ServerArgs.schedule_policy,
1047
            choices=["lpm", "random", "fcfs", "dfs-weight", "lof"],
1048
            help="The scheduling policy of the requests.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1049
        )
1050
1051
1052
1053
        parser.add_argument(
            "--schedule-conservativeness",
            type=float,
            default=ServerArgs.schedule_conservativeness,
1054
            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.",
1055
        )
1056
1057
1058
1059
1060
1061
        parser.add_argument(
            "--page-size",
            type=int,
            default=ServerArgs.page_size,
            help="The number of tokens in a page.",
        )
tarinkk's avatar
tarinkk committed
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
        parser.add_argument(
            "--hybrid-kvcache-ratio",
            nargs="?",
            const=0.5,
            type=float,
            default=ServerArgs.hybrid_kvcache_ratio,
            help=(
                "Mix ratio in [0,1] between uniform and hybrid kv buffers "
                "(0.0 = pure uniform: swa_size / full_size = 1)"
                "(1.0 = pure hybrid: swa_size / full_size = local_attention_size / context_length)"
            ),
        )
Hanming Lu's avatar
Hanming Lu committed
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
        parser.add_argument(
            "--swa-full-tokens-ratio",
            type=float,
            default=ServerArgs.swa_full_tokens_ratio,
            help="The ratio of SWA layer KV tokens / full layer KV tokens, regardless of the number of swa:full layers. It should be between 0 and 1. "
            "E.g. 0.5 means if each swa layer has 50 tokens, then each full layer has 100 tokens.",
        )
        parser.add_argument(
            "--disable-hybrid-swa-memory",
            action="store_true",
            help="Disable the hybrid SWA memory.",
        )
1086

Lianmin Zheng's avatar
Lianmin Zheng committed
1087
1088
1089
1090
1091
1092
1093
        # Runtime options
        parser.add_argument(
            "--device",
            type=str,
            default=ServerArgs.device,
            help="The device to use ('cuda', 'xpu', 'hpu', 'npu', 'cpu'). Defaults to auto-detection if not specified.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1094
        parser.add_argument(
1095
            "--tensor-parallel-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
1096
            "--tp-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
1097
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
1098
            default=ServerArgs.tp_size,
1099
            help="The tensor parallelism size.",
1100
        )
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
        parser.add_argument(
            "--pipeline-parallel-size",
            "--pp-size",
            type=int,
            default=ServerArgs.pp_size,
            help="The pipeline parallelism size.",
        )
        parser.add_argument(
            "--max-micro-batch-size",
            type=int,
            default=ServerArgs.max_micro_batch_size,
            help="The maximum micro batch size in pipeline parallelism.",
        )
1114
1115
1116
        parser.add_argument(
            "--stream-interval",
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
1117
            default=ServerArgs.stream_interval,
1118
            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",
1119
        )
1120
1121
1122
1123
1124
        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
1125
1126
1127
1128
        parser.add_argument(
            "--random-seed",
            type=int,
            default=ServerArgs.random_seed,
1129
            help="The random seed.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1130
        )
1131
1132
1133
1134
        parser.add_argument(
            "--constrained-json-whitespace-pattern",
            type=str,
            default=ServerArgs.constrained_json_whitespace_pattern,
Lianmin Zheng's avatar
Lianmin Zheng committed
1135
            help="(outlines backend only) 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 ]*",
1136
        )
1137
1138
1139
1140
1141
1142
        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.",
        )
1143
1144
1145
1146
1147
1148
        parser.add_argument(
            "--dist-timeout",
            type=int,
            default=ServerArgs.dist_timeout,
            help="Set timeout for torch.distributed initialization.",
        )
1149
1150
1151
1152
        parser.add_argument(
            "--download-dir",
            type=str,
            default=ServerArgs.download_dir,
1153
            help="Model download directory for huggingface.",
1154
        )
1155
1156
1157
1158
1159
1160
        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.",
        )
1161
1162
1163
1164
1165
1166
        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,...",
        )
1167
1168
1169
1170
1171
        parser.add_argument(
            "--sleep-on-idle",
            action="store_true",
            help="Reduce CPU usage when sglang is idle.",
        )
1172
1173

        # Logging
Lianmin Zheng's avatar
Lianmin Zheng committed
1174
1175
1176
1177
        parser.add_argument(
            "--log-level",
            type=str,
            default=ServerArgs.log_level,
1178
            help="The logging level of all loggers.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1179
        )
1180
        parser.add_argument(
1181
1182
1183
1184
            "--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.",
1185
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1186
        parser.add_argument(
1187
            "--log-requests",
Lianmin Zheng's avatar
Lianmin Zheng committed
1188
            action="store_true",
1189
1190
1191
1192
1193
            help="Log metadata, inputs, outputs of all requests. The verbosity is decided by --log-requests-level",
        )
        parser.add_argument(
            "--log-requests-level",
            type=int,
1194
            default=ServerArgs.log_requests_level,
1195
1196
1197
1198
1199
1200
1201
1202
            help="0: Log metadata (no sampling parameters). 1: Log metadata and sampling parameters. 2: Log metadata, sampling parameters and partial input/output. 3: Log every input/output.",
            choices=[0, 1, 2, 3],
        )
        parser.add_argument(
            "--crash-dump-folder",
            type=str,
            default=ServerArgs.crash_dump_folder,
            help="Folder path to dump requests from the last 5 min before a crash (if any). If not specified, crash dumping is disabled.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1203
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1204
1205
1206
        parser.add_argument(
            "--show-time-cost",
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
1207
            help="Show time cost of custom marks.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1208
        )
1209
1210
1211
1212
1213
        parser.add_argument(
            "--enable-metrics",
            action="store_true",
            help="Enable log prometheus metrics.",
        )
1214
1215
1216
1217
1218
1219
1220
        parser.add_argument(
            "--enable-metrics-for-all-schedulers",
            action="store_true",
            help="Enable --enable-metrics-for-all-schedulers when you want schedulers on all TP ranks (not just TP 0) "
            "to record request metrics separately. This is especially useful when dp_attention is enabled, as "
            "otherwise all metrics appear to come from TP 0.",
        )
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
        parser.add_argument(
            "--bucket-time-to-first-token",
            type=float,
            nargs="+",
            default=ServerArgs.bucket_time_to_first_token,
            help="The buckets of time to first token, specified as a list of floats.",
        )
        parser.add_argument(
            "--bucket-inter-token-latency",
            type=float,
            nargs="+",
            default=ServerArgs.bucket_inter_token_latency,
            help="The buckets of inter-token latency, specified as a list of floats.",
        )
        parser.add_argument(
            "--bucket-e2e-request-latency",
            type=float,
            nargs="+",
            default=ServerArgs.bucket_e2e_request_latency,
            help="The buckets of end-to-end request latency, specified as a list of floats.",
        )
        parser.add_argument(
            "--collect-tokens-histogram",
            action="store_true",
            default=ServerArgs.collect_tokens_histogram,
            help="Collect prompt/generation tokens histogram.",
        )
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
        bucket_rule = (
            "Supports 3 rule types: 'default' uses predefined buckets; 'tse <middle> <base> <count>' "
            "generates two sides exponential distributed buckets (e.g., 'tse 1000 2 8' generates buckets "
            "[984.0, 992.0, 996.0, 998.0, 1000.0, 1002.0, 1004.0, 1008.0, 1016.0]).); 'customer <value1> "
            "<value2> ...' uses custom bucket values (e.g., 'customer 10 50 100 500')."
        )
        parser.add_argument(
            "--prompt-tokens-buckets",
            type=str,
            nargs="+",
            default=ServerArgs.prompt_tokens_buckets,
            help=f"The buckets rule of prompt tokens. {bucket_rule}",
        )
        parser.add_argument(
            "--generation-tokens-buckets",
            type=str,
            nargs="+",
            default=ServerArgs.generation_tokens_buckets,
            help=f"The buckets rule for generation tokens histogram. {bucket_rule}",
        )
1268
1269
1270
1271
1272
1273
        parser.add_argument(
            "--gc-warning-threshold-secs",
            type=float,
            default=ServerArgs.gc_warning_threshold_secs,
            help="The threshold for long GC warning. If a GC takes longer than this, a warning will be logged. Set to 0 to disable.",
        )
1274
1275
1276
1277
        parser.add_argument(
            "--decode-log-interval",
            type=int,
            default=ServerArgs.decode_log_interval,
1278
            help="The log interval of decode batch.",
1279
        )
1280
1281
1282
1283
1284
1285
        parser.add_argument(
            "--enable-request-time-stats-logging",
            action="store_true",
            default=ServerArgs.enable_request_time_stats_logging,
            help="Enable per request time stats logging",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1286
1287
1288
1289
1290
1291
        parser.add_argument(
            "--kv-events-config",
            type=str,
            default=None,
            help="Config in json format for NVIDIA dynamo KV event publishing. Publishing will be enabled if this flag is used.",
        )
1292

1293
        # API related
Liangsheng Yin's avatar
Liangsheng Yin committed
1294
1295
1296
1297
        parser.add_argument(
            "--api-key",
            type=str,
            default=ServerArgs.api_key,
1298
            help="Set API key of the server. It is also used in the OpenAI API compatible server.",
Liangsheng Yin's avatar
Liangsheng Yin committed
1299
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1300
1301
1302
1303
1304
1305
        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.",
        )
1306
1307
1308
1309
1310
1311
        parser.add_argument(
            "--weight-version",
            type=str,
            default=ServerArgs.weight_version,
            help="Version identifier for the model weights. Defaults to 'default' if not specified.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
        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(
            "--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.",
        )
1324
        parser.add_argument(
1325
            "--file-storage-path",
1326
            type=str,
1327
            default=ServerArgs.file_storage_path,
1328
1329
            help="The path of the file storage in backend.",
        )
1330
1331
1332
1333
1334
        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
1335
1336
1337
1338
1339
1340
1341
        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())}.",
        )
1342
        tool_call_parser_choices = list(FunctionCallParser.ToolCallParserEnum.keys())
1343
1344
1345
        parser.add_argument(
            "--tool-call-parser",
            type=str,
1346
            choices=tool_call_parser_choices,
1347
            default=ServerArgs.tool_call_parser,
1348
            help=f"Specify the parser for handling tool-call interactions. Options include: {tool_call_parser_choices}.",
1349
        )
1350
1351
1352
1353
1354
1355
        parser.add_argument(
            "--tool-server",
            type=str,
            default=None,
            help="Either 'demo' or a comma-separated list of tool server urls to use for the model. If not specified, no tool server will be used.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1356

1357
1358
        # Data parallelism
        parser.add_argument(
1359
            "--data-parallel-size",
1360
1361
1362
            "--dp-size",
            type=int,
            default=ServerArgs.dp_size,
1363
            help="The data parallelism size.",
1364
1365
1366
1367
1368
        )
        parser.add_argument(
            "--load-balance-method",
            type=str,
            default=ServerArgs.load_balance_method,
1369
            help="The load balancing strategy for data parallelism.",
1370
1371
1372
            choices=[
                "round_robin",
                "shortest_queue",
1373
                "minimum_tokens",
1374
1375
            ],
        )
1376

1377
        # Multi-node distributed serving
1378
        parser.add_argument(
1379
            "--dist-init-addr",
1380
            "--nccl-init-addr",  # For backward compatibility. This will be removed in the future.
1381
            type=str,
1382
            help="The host address for initializing distributed backend (e.g., `192.168.0.2:25000`).",
1383
1384
        )
        parser.add_argument(
Liangsheng Yin's avatar
Liangsheng Yin committed
1385
            "--nnodes", type=int, default=ServerArgs.nnodes, help="The number of nodes."
1386
        )
1387
1388
1389
        parser.add_argument(
            "--node-rank", type=int, default=ServerArgs.node_rank, help="The node rank."
        )
1390

Lianmin Zheng's avatar
Lianmin Zheng committed
1391
1392
1393
1394
1395
1396
1397
        # 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,
        )
1398
1399
1400
1401
1402
        parser.add_argument(
            "--preferred-sampling-params",
            type=str,
            help="json-formatted sampling settings that will be returned in /get_model_info",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1403

1404
        # LoRA
1405
1406
1407
1408
1409
1410
        parser.add_argument(
            "--enable-lora",
            default=ServerArgs.enable_lora,
            action="store_true",
            help="Enable LoRA support for the model. This argument is automatically set to True if `--lora-paths` is provided for backward compatibility.",
        )
1411
1412
1413
1414
1415
1416
1417
1418
1419
        parser.add_argument(
            "--max-lora-rank",
            default=ServerArgs.max_lora_rank,
            type=int,
            help="The maximum rank of LoRA adapters. If not specified, it will be automatically inferred from the adapters provided in --lora-paths.",
        )
        parser.add_argument(
            "--lora-target-modules",
            type=str,
1420
            choices=SUPPORTED_LORA_TARGET_MODULES + [LORA_TARGET_ALL_MODULES],
1421
1422
            nargs="*",
            default=None,
1423
1424
1425
            help="The union set of all target modules where LoRA should be applied. If not specified, "
            "it will be automatically inferred from the adapters provided in --lora-paths. If 'all' is specified, "
            "all supported modules will be targeted.",
1426
        )
1427
1428
1429
1430
1431
1432
        parser.add_argument(
            "--lora-paths",
            type=str,
            nargs="*",
            default=None,
            action=LoRAPathAction,
1433
            help='The list of LoRA adapters to load. Each adapter must be specified in one of the following formats: <PATH> | <NAME>=<PATH> | JSON with schema {"lora_name":str,"lora_path":str,"pinned":bool}',
1434
1435
1436
1437
1438
        )
        parser.add_argument(
            "--max-loras-per-batch",
            type=int,
            default=8,
1439
1440
            help="Maximum number of adapters for a running batch, include base-only request.",
        )
1441
1442
1443
1444
1445
1446
        parser.add_argument(
            "--max-loaded-loras",
            type=int,
            default=ServerArgs.max_loaded_loras,
            help="If specified, it limits the maximum number of LoRA adapters loaded in CPU memory at a time. The value must be greater than or equal to `--max-loras-per-batch`.",
        )
1447
1448
1449
1450
1451
        parser.add_argument(
            "--lora-backend",
            type=str,
            default="triton",
            help="Choose the kernel backend for multi-LoRA serving.",
1452
1453
1454
        )

        # Kernel backend
1455
1456
1457
        parser.add_argument(
            "--attention-backend",
            type=str,
1458
            choices=ATTENTION_BACKEND_CHOICES,
1459
1460
1461
            default=ServerArgs.attention_backend,
            help="Choose the kernels for attention layers.",
        )
1462
1463
1464
        parser.add_argument(
            "--prefill-attention-backend",
            type=str,
1465
            choices=ATTENTION_BACKEND_CHOICES,
1466
1467
1468
            default=ServerArgs.prefill_attention_backend,
            help="Choose the kernels for prefill attention layers (have priority over --attention-backend).",
        )
1469
1470
1471
        parser.add_argument(
            "--decode-attention-backend",
            type=str,
1472
            choices=ATTENTION_BACKEND_CHOICES,
1473
1474
1475
            default=ServerArgs.decode_attention_backend,
            help="Choose the kernels for decode attention layers (have priority over --attention-backend).",
        )
1476
1477
1478
1479
1480
1481
1482
        parser.add_argument(
            "--sampling-backend",
            type=str,
            choices=["flashinfer", "pytorch"],
            default=ServerArgs.sampling_backend,
            help="Choose the kernels for sampling layers.",
        )
1483
1484
1485
        parser.add_argument(
            "--grammar-backend",
            type=str,
1486
            choices=GRAMMAR_BACKEND_CHOICES,
1487
            default=ServerArgs.grammar_backend,
Lianmin Zheng's avatar
Lianmin Zheng committed
1488
            help="Choose the backend for grammar-guided decoding.",
1489
        )
1490
1491
1492
1493
1494
1495
1496
        parser.add_argument(
            "--mm-attention-backend",
            type=str,
            choices=["sdpa", "fa3", "triton_attn"],
            default=ServerArgs.mm_attention_backend,
            help="Set multimodal attention backend.",
        )
1497

1498
1499
1500
1501
        # Speculative decoding
        parser.add_argument(
            "--speculative-algorithm",
            type=str,
James Liu's avatar
James Liu committed
1502
            choices=["EAGLE", "EAGLE3", "NEXTN"],
1503
1504
1505
1506
1507
1508
1509
            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.",
        )
1510
1511
1512
1513
1514
1515
1516
1517
        parser.add_argument(
            "--speculative-draft-model-revision",
            type=str,
            default=None,
            help="The specific draft model version to use. It can be a branch "
            "name, a tag name, or a commit id. If unspecified, will use "
            "the default version.",
        )
1518
1519
1520
1521
1522
1523
1524
1525
1526
        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,
1527
            help="The number of tokens sampled from the draft model in eagle2 each step.",
1528
1529
            default=ServerArgs.speculative_eagle_topk,
        )
1530
1531
1532
        parser.add_argument(
            "--speculative-num-draft-tokens",
            type=int,
1533
            help="The number of tokens sampled from the draft model in Speculative Decoding.",
1534
1535
            default=ServerArgs.speculative_num_draft_tokens,
        )
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
        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,
        )
1548
1549
1550
1551
1552
1553
        parser.add_argument(
            "--speculative-token-map",
            type=str,
            help="The path of the draft model's small vocab table.",
            default=ServerArgs.speculative_token_map,
        )
1554
1555
1556
1557
1558

        # Expert parallelism
        parser.add_argument(
            "--expert-parallel-size",
            "--ep-size",
Cheng Wan's avatar
Cheng Wan committed
1559
            "--ep",
1560
1561
1562
1563
1564
            type=int,
            default=ServerArgs.ep_size,
            help="The expert parallelism size.",
        )
        parser.add_argument(
1565
1566
            "--moe-a2a-backend",
            type=str,
1567
            choices=["none", "deepep"],
1568
1569
            default=ServerArgs.moe_a2a_backend,
            help="Choose the backend for MoE A2A.",
1570
        )
1571
        parser.add_argument(
1572
1573
1574
1575
1576
1577
1578
1579
            "--moe-runner-backend",
            type=str,
            choices=[
                "auto",
                "triton",
                "triton_kernel",
                "flashinfer_trtllm",
                "flashinfer_cutlass",
1580
                "flashinfer_mxfp4",
1581
1582
1583
            ],
            default=ServerArgs.moe_runner_backend,
            help="Choose the runner backend for MoE.",
1584
1585
        )
        parser.add_argument(
1586
1587
1588
1589
1590
1591
1592
            "--flashinfer-mxfp4-moe-precision",
            type=str,
            choices=["mxfp4", "bf16"],
            default=ServerArgs.flashinfer_mxfp4_moe_precision,
            help="Choose the computation precision of flashinfer mxfp4 moe",
        )
        parser.add_argument(
1593
1594
            "--enable-flashinfer-allreduce-fusion",
            action="store_true",
1595
            help="Enable FlashInfer allreduce fusion with Residual RMSNorm.",
1596
        )
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
        parser.add_argument(
            "--deepep-mode",
            type=str,
            choices=["normal", "low_latency", "auto"],
            default="auto",
            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.",
        )
        parser.add_argument(
            "--ep-num-redundant-experts",
            type=int,
            default=ServerArgs.ep_num_redundant_experts,
            help="Allocate this number of redundant experts in expert parallel.",
        )
        parser.add_argument(
            "--ep-dispatch-algorithm",
            type=str,
            default=ServerArgs.ep_dispatch_algorithm,
            help="The algorithm to choose ranks for redundant experts in expert parallel.",
        )
        parser.add_argument(
            "--init-expert-location",
            type=str,
            default=ServerArgs.init_expert_location,
            help="Initial location of EP experts.",
        )
        parser.add_argument(
            "--enable-eplb",
            action="store_true",
            help="Enable EPLB algorithm",
        )
        parser.add_argument(
            "--eplb-algorithm",
            type=str,
            default=ServerArgs.eplb_algorithm,
            help="Chosen EPLB algorithm",
        )
        parser.add_argument(
            "--eplb-rebalance-num-iterations",
            type=int,
            default=ServerArgs.eplb_rebalance_num_iterations,
            help="Number of iterations to automatically trigger a EPLB re-balance.",
        )
        parser.add_argument(
            "--eplb-rebalance-layers-per-chunk",
            type=int,
            default=ServerArgs.eplb_rebalance_layers_per_chunk,
            help="Number of layers to rebalance per forward pass.",
        )
1645
1646
1647
1648
1649
1650
        parser.add_argument(
            "--eplb-min-rebalancing-utilization-threshold",
            type=float,
            default=ServerArgs.eplb_min_rebalancing_utilization_threshold,
            help="Minimum threshold for GPU average utilization to trigger EPLB rebalancing. Must be in the range [0.0, 1.0].",
        )
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
        parser.add_argument(
            "--expert-distribution-recorder-mode",
            type=str,
            default=ServerArgs.expert_distribution_recorder_mode,
            help="Mode of expert distribution recorder.",
        )
        parser.add_argument(
            "--expert-distribution-recorder-buffer-size",
            type=int,
            default=ServerArgs.expert_distribution_recorder_buffer_size,
            help="Circular buffer size of expert distribution recorder. Set to -1 to denote infinite buffer.",
        )
        parser.add_argument(
            "--enable-expert-distribution-metrics",
            action="store_true",
            help="Enable logging metrics for expert balancedness",
        )
        parser.add_argument(
            "--deepep-config",
            type=str,
            default=ServerArgs.deepep_config,
            help="Tuned DeepEP config suitable for your own cluster. It can be either a string with JSON content or a file path.",
        )
        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.",
        )
1680

Lianmin Zheng's avatar
Lianmin Zheng committed
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
        # Hierarchical cache
        parser.add_argument(
            "--enable-hierarchical-cache",
            action="store_true",
            help="Enable hierarchical cache",
        )
        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.",
        )
        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.",
        )
        parser.add_argument(
            "--hicache-io-backend",
            type=str,
            choices=["direct", "kernel"],
            default=ServerArgs.hicache_io_backend,
            help="The IO backend for KV cache transfer between CPU and GPU",
        )
1713
1714
1715
1716
1717
1718
1719
        parser.add_argument(
            "--hicache-mem-layout",
            type=str,
            choices=["layer_first", "page_first"],
            default=ServerArgs.hicache_mem_layout,
            help="The layout of host memory pool for hierarchical cache.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1720
1721
1722
        parser.add_argument(
            "--hicache-storage-backend",
            type=str,
1723
            choices=["file", "mooncake", "hf3fs", "nixl"],
Lianmin Zheng's avatar
Lianmin Zheng committed
1724
1725
1726
            default=ServerArgs.hicache_storage_backend,
            help="The storage backend for hierarchical KV cache.",
        )
pansicheng's avatar
pansicheng committed
1727
1728
1729
1730
1731
1732
1733
        parser.add_argument(
            "--hicache-storage-prefetch-policy",
            type=str,
            choices=["best_effort", "wait_complete", "timeout"],
            default=ServerArgs.hicache_storage_prefetch_policy,
            help="Control when prefetching from the storage backend should stop.",
        )
1734
1735
1736
1737
1738
1739
        parser.add_argument(
            "--hicache-storage-backend-extra-config",
            type=str,
            default=ServerArgs.hicache_storage_backend_extra_config,
            help="A dictionary in JSON string format containing extra configuration for the storage backend.",
        )
1740
1741
1742
1743
1744
1745
        # LMCache
        parser.add_argument(
            "--enable-lmcache",
            action="store_true",
            help="Using LMCache as an alternative hierarchical cache solution",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1746

1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
        # 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",
        )

fzyzcjy's avatar
fzyzcjy committed
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
        # Offloading
        parser.add_argument(
            "--cpu-offload-gb",
            type=int,
            default=ServerArgs.cpu_offload_gb,
            help="How many GBs of RAM to reserve for CPU offloading.",
        )
        parser.add_argument(
            "--offload-group-size",
            type=int,
            default=ServerArgs.offload_group_size,
            help="Number of layers per group in offloading.",
        )
        parser.add_argument(
            "--offload-num-in-group",
            type=int,
            default=ServerArgs.offload_num_in_group,
            help="Number of layers to be offloaded within a group.",
        )
        parser.add_argument(
            "--offload-prefetch-step",
            type=int,
            default=ServerArgs.offload_prefetch_step,
            help="Steps to prefetch in offloading.",
        )
        parser.add_argument(
            "--offload-mode",
            type=str,
            default=ServerArgs.offload_mode,
            help="Mode of offloading.",
        )

1816
        # Optimization/debug options
Liangsheng Yin's avatar
Liangsheng Yin committed
1817
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
1818
            "--disable-radix-cache",
Liangsheng Yin's avatar
Liangsheng Yin committed
1819
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
1820
            help="Disable RadixAttention for prefix caching.",
Liangsheng Yin's avatar
Liangsheng Yin committed
1821
        )
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
        parser.add_argument(
            "--cuda-graph-max-bs",
            type=int,
            default=ServerArgs.cuda_graph_max_bs,
            help="Set the maximum batch size for cuda graph. It will extend the cuda graph capture batch size to this value.",
        )
        parser.add_argument(
            "--cuda-graph-bs",
            type=int,
            nargs="+",
            help="Set the list of batch sizes for cuda graph.",
        )
1834
1835
1836
        parser.add_argument(
            "--disable-cuda-graph",
            action="store_true",
1837
            help="Disable cuda graph.",
1838
        )
1839
        parser.add_argument(
1840
1841
            "--disable-cuda-graph-padding",
            action="store_true",
1842
            help="Disable cuda graph when padding is needed. Still uses cuda graph when padding is not needed.",
1843
        )
1844
1845
1846
1847
1848
        parser.add_argument(
            "--enable-profile-cuda-graph",
            action="store_true",
            help="Enable profiling of cuda graph capture.",
        )
1849
1850
1851
1852
1853
        parser.add_argument(
            "--enable-cudagraph-gc",
            action="store_true",
            help="Enable garbage collection during CUDA graph capture. If disabled (default), GC is frozen during capture to speed up the process.",
        )
1854
1855
1856
1857
1858
        parser.add_argument(
            "--enable-nccl-nvls",
            action="store_true",
            help="Enable NCCL NVLS for prefill heavy requests when available.",
        )
1859
1860
1861
1862
1863
        parser.add_argument(
            "--enable-symm-mem",
            action="store_true",
            help="Enable NCCL symmetric memory for fast collectives.",
        )
1864
1865
1866
1867
1868
        parser.add_argument(
            "--disable-flashinfer-cutlass-moe-fp4-allgather",
            action="store_true",
            help="Disables quantize before all-gather for flashinfer cutlass moe.",
        )
1869
1870
1871
1872
1873
        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.",
        )
1874
        parser.add_argument(
1875
            "--disable-outlines-disk-cache",
1876
            action="store_true",
1877
            help="Disable disk cache of outlines to avoid possible crashes related to file system or high concurrency.",
1878
        )
1879
1880
1881
1882
1883
        parser.add_argument(
            "--disable-custom-all-reduce",
            action="store_true",
            help="Disable the custom all-reduce kernel and fall back to NCCL.",
        )
1884
1885
1886
1887
1888
        parser.add_argument(
            "--enable-mscclpp",
            action="store_true",
            help="Enable using mscclpp for small messages for all-reduce kernel and fall back to NCCL.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1889
        parser.add_argument(
1890
            "--disable-overlap-schedule",
Lianmin Zheng's avatar
Lianmin Zheng committed
1891
            action="store_true",
1892
            help="Disable the overlap scheduler, which overlaps the CPU scheduler with GPU model worker.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1893
        )
1894
1895
1896
        parser.add_argument(
            "--enable-mixed-chunk",
            action="store_true",
1897
            help="Enabling mixing prefill and decode in a batch when using chunked prefill.",
1898
        )
Ke Bao's avatar
Ke Bao committed
1899
1900
1901
        parser.add_argument(
            "--enable-dp-attention",
            action="store_true",
1902
            help="Enabling data parallelism for attention and tensor parallelism for FFN. The dp size should be equal to the tp size. Currently DeepSeek-V2 and Qwen 2/3 MoE models are supported.",
Ke Bao's avatar
Ke Bao committed
1903
        )
1904
1905
1906
1907
1908
        parser.add_argument(
            "--enable-dp-lm-head",
            action="store_true",
            help="Enable vocabulary parallel across the attention TP group to avoid all-gather across DP groups, optimizing performance under DP attention.",
        )
1909
1910
1911
1912
1913
        parser.add_argument(
            "--enable-two-batch-overlap",
            action="store_true",
            help="Enabling two micro batches to overlap.",
        )
1914
1915
1916
1917
1918
1919
        parser.add_argument(
            "--tbo-token-distribution-threshold",
            type=float,
            default=ServerArgs.tbo_token_distribution_threshold,
            help="The threshold of token distribution between two batches in micro-batch-overlap, determines whether to two-batch-overlap or two-chunk-overlap. Set to 0 denote disable two-chunk-overlap.",
        )
1920
1921
1922
        parser.add_argument(
            "--enable-torch-compile",
            action="store_true",
1923
1924
            help="Optimize the model with torch.compile. Experimental feature.",
        )
1925
        parser.add_argument(
1926
            "--torch-compile-max-bs",
1927
            type=int,
1928
            default=ServerArgs.torch_compile_max_bs,
1929
1930
            help="Set the maximum batch size when using torch compile.",
        )
1931
1932
1933
1934
        parser.add_argument(
            "--torchao-config",
            type=str,
            default=ServerArgs.torchao_config,
1935
            help="Optimize the model with torchao. Experimental feature. Current choices are: int8dq, int8wo, int4wo-<group_size>, fp8wo, fp8dq-per_tensor, fp8dq-per_row",
1936
        )
1937
1938
1939
1940
1941
        parser.add_argument(
            "--enable-nan-detection",
            action="store_true",
            help="Enable the NaN detection for debugging purposes.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1942
        parser.add_argument(
1943
            "--enable-p2p-check",
Lianmin Zheng's avatar
Lianmin Zheng committed
1944
            action="store_true",
1945
            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1946
        )
1947
        parser.add_argument(
1948
            "--triton-attention-reduce-in-fp32",
1949
            action="store_true",
1950
            help="Cast the intermediate attention results to fp32 to avoid possible crashes related to fp16."
1951
            "This only affects Triton attention kernels.",
1952
        )
1953
1954
1955
1956
1957
1958
        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.",
        )
1959
1960
1961
1962
1963
1964
1965
1966
        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.",
        )
1967
1968
1969
1970
1971
        parser.add_argument(
            "--delete-ckpt-after-loading",
            action="store_true",
            help="Delete the model checkpoint after loading the model.",
        )
1972
1973
1974
1975
1976
        parser.add_argument(
            "--enable-memory-saver",
            action="store_true",
            help="Allow saving memory using release_memory_occupation and resume_memory_occupation",
        )
1977
1978
1979
1980
1981
        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.",
        )
1982
1983
1984
1985
1986
        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)",
        )
1987
        parser.add_argument(
1988
            "--flashinfer-mla-disable-ragged",
1989
            action="store_true",
1990
            help="Not using ragged prefill wrapper when running flashinfer mla",
1991
        )
1992
        parser.add_argument(
1993
1994
1995
            "--disable-shared-experts-fusion",
            action="store_true",
            help="Disable shared experts fusion optimization for deepseek v3/r1.",
1996
        )
1997
1998
1999
2000
2001
        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
2002
2003
2004
2005
2006
        parser.add_argument(
            "--disable-fast-image-processor",
            action="store_true",
            help="Adopt base image processor instead of fast image processor.",
        )
2007
2008
2009
2010
2011
        parser.add_argument(
            "--enable-return-hidden-states",
            action="store_true",
            help="Enable returning hidden states with responses.",
        )
2012
2013
2014
2015
2016
2017
        parser.add_argument(
            "--scheduler-recv-interval",
            type=int,
            default=ServerArgs.scheduler_recv_interval,
            help="The interval to poll requests in scheduler. Can be set to >1 to reduce the overhead of this.",
        )
2018
2019
2020
2021
2022
2023
        parser.add_argument(
            "--numa-node",
            type=int,
            nargs="+",
            help="Sets the numa node for the subprocesses. i-th element corresponds to i-th subprocess.",
        )
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043

        # 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.",
        )
2044
2045
2046
2047
2048
        parser.add_argument(
            "--debug-tensor-dump-prefill-only",
            action="store_true",
            help="Only dump the tensors for prefill requests (i.e. batch size > 1).",
        )
2049

Lianmin Zheng's avatar
Lianmin Zheng committed
2050
        # PD disaggregation
Byron Hsu's avatar
Byron Hsu committed
2051
2052
2053
2054
2055
2056
2057
        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',
        )
2058
2059
2060
2061
        parser.add_argument(
            "--disaggregation-transfer-backend",
            type=str,
            default=ServerArgs.disaggregation_transfer_backend,
2062
            choices=DISAGG_TRANSFER_BACKEND_CHOICES,
2063
2064
            help="The backend for disaggregation transfer. Default is mooncake.",
        )
2065
2066
2067
2068
2069
2070
        parser.add_argument(
            "--disaggregation-bootstrap-port",
            type=int,
            default=ServerArgs.disaggregation_bootstrap_port,
            help="Bootstrap server port on the prefill server. Default is 8998.",
        )
Byron Hsu's avatar
Byron Hsu committed
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
        parser.add_argument(
            "--disaggregation-decode-tp",
            type=int,
            default=ServerArgs.disaggregation_decode_tp,
            help="Decode tp size. If not set, it matches the tp size of the current engine. This is only set on the prefill server.",
        )
        parser.add_argument(
            "--disaggregation-decode-dp",
            type=int,
            default=ServerArgs.disaggregation_decode_dp,
            help="Decode dp size. If not set, it matches the dp size of the current engine. This is only set on the prefill server.",
        )
        parser.add_argument(
            "--disaggregation-prefill-pp",
            type=int,
            default=ServerArgs.disaggregation_prefill_pp,
            help="Prefill pp size. If not set, it is default to 1. This is only set on the decode server.",
        )
2089
2090
2091
2092
        parser.add_argument(
            "--disaggregation-ib-device",
            type=str,
            default=ServerArgs.disaggregation_ib_device,
2093
2094
2095
            help="The InfiniBand devices for disaggregation transfer, accepts single device (e.g., --disaggregation-ib-device mlx5_0) "
            "or multiple comma-separated devices (e.g., --disaggregation-ib-device mlx5_0,mlx5_1). "
            "Default is None, which triggers automatic device detection when mooncake backend is enabled.",
2096
        )
2097
2098
2099
2100
2101
2102
        parser.add_argument(
            "--num-reserved-decode-tokens",
            type=int,
            default=ServerArgs.num_reserved_decode_tokens,
            help="Number of decode tokens that will have memory reserved when adding new request to the running batch.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2103
2104

        # Custom weight loader
2105
2106
2107
2108
2109
2110
2111
        parser.add_argument(
            "--custom-weight-loader",
            type=str,
            nargs="*",
            default=None,
            help="The custom dataloader which used to update the model. Should be set with a valid import path, such as my_package.weight_load_func",
        )
2112
2113
2114
2115
2116
2117
2118
        parser.add_argument(
            "--weight-loader-disable-mmap",
            action="store_true",
            help="Disable mmap while loading weight using safetensors.",
        )

        # For PD-Multiplexing
2119
2120
2121
2122
2123
        parser.add_argument(
            "--enable-pdmux",
            action="store_true",
            help="Enable PD-Multiplexing, PD running on greenctx stream.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2124

2125
2126
2127
2128
2129
2130
        parser.add_argument(
            "--sm-group-num",
            type=int,
            default=ServerArgs.sm_group_num,
            help="Number of sm partition groups.",
        )
Byron Hsu's avatar
Byron Hsu committed
2131

2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
        # Deprecated arguments
        parser.add_argument(
            "--enable-ep-moe",
            action="store_true",
            help="(Deprecated) Enabling expert parallelism for moe. The ep size is equal to the tp size.",
        )
        parser.add_argument(
            "--enable-deepep-moe",
            action="store_true",
            help="(Deprecated) Enabling DeepEP MoE implementation for EP MoE.",
        )
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
        parser.add_argument(
            "--enable-flashinfer-cutlass-moe",
            action="store_true",
            help="(Deprecated) Enable FlashInfer CUTLASS MoE backend for modelopt_fp4 quant on Blackwell. Supports MoE-EP",
        )
        parser.add_argument(
            "--enable-flashinfer-trtllm-moe",
            action="store_true",
            help="(Deprecated) Enable FlashInfer TRTLLM MoE backend on Blackwell. Supports BlockScale FP8 MoE-EP",
        )
        parser.add_argument(
            "--enable-triton-kernel-moe",
            action="store_true",
            help="(Deprecated) Use triton moe grouped gemm kernel.",
        )
2158
2159
2160
2161
2162
        parser.add_argument(
            "--enable-flashinfer-mxfp4-moe",
            action="store_true",
            help="(Deprecated) Enable FlashInfer MXFP4 MoE backend for modelopt_fp4 quant on Blackwell.",
        )
2163

Lianmin Zheng's avatar
Lianmin Zheng committed
2164
2165
    @classmethod
    def from_cli_args(cls, args: argparse.Namespace):
2166
        args.tp_size = args.tensor_parallel_size
2167
        args.pp_size = args.pipeline_parallel_size
2168
        args.dp_size = args.data_parallel_size
xiaobochen's avatar
xiaobochen committed
2169
        args.ep_size = args.expert_parallel_size
Lianmin Zheng's avatar
Lianmin Zheng committed
2170
2171
2172
2173
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        return cls(**{attr: getattr(args, attr) for attr in attrs})

    def url(self):
2174
        if is_valid_ipv6_address(self.host):
2175
2176
2177
            return f"http://[{self.host}]:{self.port}"
        else:
            return f"http://{self.host}:{self.port}"
Lianmin Zheng's avatar
Lianmin Zheng committed
2178

Lianmin Zheng's avatar
Lianmin Zheng committed
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
    def get_hf_config(self):
        kwargs = {}
        hf_config = get_config(
            self.model_path,
            trust_remote_code=self.trust_remote_code,
            revision=self.revision,
            model_override_args=json.loads(self.json_model_override_args),
            **kwargs,
        )
        return hf_config

2190
    def check_server_args(self):
2191
        # Check parallel size constraints
2192
        assert (
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
            self.tp_size * self.pp_size
        ) % self.nnodes == 0, "tp_size must be divisible by number of nodes"

        if self.pp_size > 1:
            assert (
                self.disable_overlap_schedule
                and self.speculative_algorithm is None
                and not self.enable_mixed_chunk
            ), "Pipeline parallelism is not compatible with overlap schedule, speculative decoding, mixed chunked prefill."

2203
        assert not (
2204
2205
            self.dp_size > 1 and self.nnodes != 1 and not self.enable_dp_attention
        ), "multi-node data parallel is not supported unless dp attention!"
2206

2207
        assert self.base_gpu_id >= 0, "base_gpu_id must be non-negative"
2208
        assert self.gpu_id_step >= 1, "gpu_id_step must be positive"
2209

Lianmin Zheng's avatar
Lianmin Zheng committed
2210
2211
2212
2213
2214
        assert self.moe_dense_tp_size in {
            1,
            None,
        }, "moe_dense_tp_size only support 1 and None currently"

2215
        # Check LoRA
2216
2217
        self.check_lora_server_args()

2218
2219
2220
2221
2222
2223
2224
        # Check speculative decoding
        if self.speculative_algorithm is not None:
            assert (
                not self.enable_mixed_chunk
            ), "enable_mixed_chunk is required for speculative decoding"

        # Check chunked prefill
2225
2226
2227
2228
2229
        # Skip validation if chunked prefill is disabled (i.e., size <= 0).
        if self.chunked_prefill_size > 0:
            assert (
                self.chunked_prefill_size % self.page_size == 0
            ), "chunked_prefill_size must be divisible by page_size"
2230

2231
2232
        # Check multi tokenizer
        assert self.tokenizer_worker_num > 0, "Tokenizer worker num must >= 1"
2233
2234
2235
2236
2237
2238
        self.validate_buckets_rule(
            "--prompt-tokens-buckets", self.prompt_tokens_buckets
        )
        self.validate_buckets_rule(
            "--generation-tokens-buckets", self.generation_tokens_buckets
        )
2239

2240
    def check_lora_server_args(self):
2241
        assert self.max_loras_per_batch > 0, "max_loras_per_batch must be positive"
2242

2243
2244
2245
2246
        # Enable LoRA if any LoRA paths are provided for backward compatibility.
        if self.lora_paths:
            if self.enable_lora is None:
                self.enable_lora = True
2247
                logger.warning(
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
                    "--enable-lora is set to True because --lora-paths is provided."
                )
            elif self.enable_lora is False:
                logger.warning(
                    "--enable-lora is set to False, any provided lora_paths will be ignored."
                )

        if self.enable_lora:
            if isinstance(self.lora_paths, list):
                lora_paths = self.lora_paths
2258
                self.lora_paths = []
2259
                for lora_path in lora_paths:
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
                    if isinstance(lora_path, str):
                        if "=" in lora_path:
                            name, path = lora_path.split("=", 1)
                            lora_ref = LoRARef(
                                lora_name=name, lora_path=path, pinned=False
                            )
                        else:
                            lora_ref = LoRARef(
                                lora_name=lora_path, lora_path=lora_path, pinned=False
                            )
                    elif isinstance(lora_path, dict):
                        assert (
                            "lora_name" in lora_path and "lora_path" in lora_path
                        ), f"When providing LoRA paths as a list of dict, each dict should contain 'lora_name' and 'lora_path' keys. Got: {lora_path}"
                        lora_ref = LoRARef(
                            lora_name=lora_path["lora_name"],
                            lora_path=lora_path["lora_path"],
                            pinned=lora_path.get("pinned", False),
2278
                        )
2279
                    else:
2280
2281
2282
                        raise ValueError(
                            f"Invalid type for item in --lora-paths list: {type(lora_path)}. "
                            "Expected a string or a dictionary."
2283
                        )
2284
                    self.lora_paths.append(lora_ref)
2285
            elif isinstance(self.lora_paths, dict):
2286
2287
                self.lora_paths = [
                    LoRARef(lora_name=k, lora_path=v, pinned=False)
2288
                    for k, v in self.lora_paths.items()
2289
                ]
2290
            elif self.lora_paths is None:
2291
                self.lora_paths = []
2292
2293
2294
2295
2296
            else:
                raise ValueError(
                    f"Invalid type for --lora-paths: {type(self.lora_paths)}. "
                    "Expected a list or a dictionary."
                )
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310

            # Expand target modules
            if self.lora_target_modules:
                self.lora_target_modules = set(self.lora_target_modules)
                if "all" in self.lora_target_modules:
                    assert (
                        len(self.lora_target_modules) == 1
                    ), "If 'all' is specified in --lora-target-modules, it should be the only module specified."
                    self.lora_target_modules = set(SUPPORTED_LORA_TARGET_MODULES)

            # Ensure sufficient information is provided for LoRA initialization.
            assert self.lora_paths or (
                self.max_lora_rank and self.lora_target_modules
            ), "When no initial --lora-paths is provided, you need to specify both --max-lora-rank and --lora-target-modules for LoRA initialization."
2311

2312
2313
2314
2315
2316
2317
            # Validate max_loaded_loras
            if self.max_loaded_loras is not None:
                assert self.max_loaded_loras >= self.max_loras_per_batch, (
                    "max_loaded_loras should be greater than or equal to max_loras_per_batch. "
                    f"max_loaded_loras={self.max_loaded_loras}, max_loras_per_batch={self.max_loras_per_batch}"
                )
2318
                assert len(self.lora_paths) <= self.max_loaded_loras, (
2319
2320
2321
2322
                    "The number of LoRA paths should not exceed max_loaded_loras. "
                    f"max_loaded_loras={self.max_loaded_loras}, lora_paths={len(self.lora_paths)}"
                )

Lianmin Zheng's avatar
Lianmin Zheng committed
2323
2324
2325
2326
2327
2328
2329
2330
    def validate_disagg_tp_size(self, prefill_tp: int, decode_tp: int):
        larger_tp = max(decode_tp, prefill_tp)
        smaller_tp = min(decode_tp, prefill_tp)
        assert larger_tp % smaller_tp == 0, (
            "Different tp size is supported only when one tp is multiple of the other. "
            f"decode_tp={decode_tp}, prefill_tp={prefill_tp}"
        )

2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
    def validate_buckets_rule(self, arg_name: str, buckets_rule: List[str]):
        if not buckets_rule:
            return

        assert len(buckets_rule) > 0, f"{arg_name} cannot be empty list"
        rule = buckets_rule[0]
        assert rule in [
            "tse",
            "default",
            "customer",
        ], f"Unsupported {arg_name} rule type: '{rule}'. Must be one of: 'tse', 'default', 'customer'"

        if rule == "tse":
            assert (
                len(buckets_rule) == 4
            ), f"{arg_name} TSE rule requires exactly 4 parameters: ['tse', middle, base, count], got {len(buckets_rule)}"
            try:
                middle = float(buckets_rule[1])
                base = float(buckets_rule[2])
                count = int(buckets_rule[3])
            except (ValueError, IndexError):
                assert (
                    False
                ), f"{arg_name} TSE rule parameters must be: ['tse', <float:middle>, <float:base>, <int:count>]"
            assert base > 1, f"{arg_name} TSE base must be larger than 1, got: {base}"
            assert count > 0, f"{arg_name} TSE count must be positive, got: {count}"
            assert middle > 0, f"{arg_name} TSE middle must be positive, got: {middle}"

        elif rule == "default":
            assert (
                len(buckets_rule) == 1
            ), f"{arg_name} default rule should only have one parameter: ['default'], got {len(buckets_rule)}"

        elif rule == "customer":
            assert (
                len(buckets_rule) >= 2
            ), f"{arg_name} customer rule requires at least one bucket value: ['customer', value1, ...]"
            try:
                bucket_values = [float(x) for x in buckets_rule[1:]]
            except ValueError:
                assert False, f"{arg_name} customer rule bucket values must be numeric"
            assert len(set(bucket_values)) == len(
                bucket_values
            ), f"{arg_name} customer rule bucket values should not contain duplicates"
            assert all(
                val >= 0 for val in bucket_values
            ), f"{arg_name} customer rule bucket values should be non-negative"

2379
2380
2381
2382
2383
    def model_specific_adjustments(self):
        hf_config = self.get_hf_config()
        model_arch = hf_config.architectures[0]
        if model_arch in ["GptOssForCausalLM"]:
            if self.attention_backend is None:
2384
                if is_cuda() and is_sm100_supported():
2385
                    self.attention_backend = "trtllm_mha"
2386
                elif is_cuda() and is_sm90_supported():
2387
2388
2389
                    self.attention_backend = "fa3"
                else:
                    self.attention_backend = "triton"
2390
            supported_backends = ["triton", "trtllm_mha", "fa3"]
2391
2392
2393
            logger.info(
                f"Use {self.attention_backend} as attention backend for GptOssForCausalLM"
            )
2394
2395
2396
            assert (
                self.attention_backend in supported_backends
            ), f"GptOssForCausalLM requires one of {supported_backends} attention backend, but got '{self.attention_backend}'"
2397
2398

            if is_sm100_supported():
2399
2400
2401
2402
2403
                if not self.enable_dp_attention:
                    self.enable_flashinfer_allreduce_fusion = True
                    logger.info(
                        "Enable FlashInfer AllReduce Fusion on sm100 for GptOssForCausalLM"
                    )
2404
2405
2406
2407
2408
2409
2410
            quantization_config = getattr(hf_config, "quantization_config", None)
            is_mxfp4_quant_format = (
                quantization_config is not None
                and quantization_config.get("quant_method") == "mxfp4"
            )

            if is_sm100_supported() and is_mxfp4_quant_format:
2411
                self.moe_runner_backend = "flashinfer_mxfp4"
2412
2413
2414
2415
                logger.warning(
                    "Detected SM100 and MXFP4 quantization format for GPT-OSS model, enabling FlashInfer MXFP4 MOE kernel."
                )
            else:
2416
                if self.moe_runner_backend == "triton_kernel":
2417
2418
2419
                    assert (
                        self.ep_size == 1
                    ), "Triton kernel MoE is only supported when ep_size == 1"
2420
2421
2422
2423
2424
2425
                if (
                    self.moe_runner_backend == "auto"
                    and self.ep_size == 1
                    and is_triton_kernels_available()
                ):
                    self.moe_runner_backend = "triton_kernel"
2426
2427
2428
2429
2430
2431
2432
                    logger.warning(
                        "Detected GPT-OSS model, enabling triton_kernels MOE kernel."
                    )
            self.disable_hybrid_swa_memory = True
            if is_mxfp4_quant_format:
                # use bf16 for mxfp4 triton kernels
                self.dtype = "bfloat16"
2433

2434
        elif "Llama4" in model_arch:
2435
2436
2437
            assert self.attention_backend in {
                "fa3",
                "aiter",
2438
2439
                "triton",
            }, "fa3, aiter, or triton is required for Llama4 model"
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
        elif model_arch in [
            "Gemma2ForCausalLM",
            "Gemma3ForCausalLM",
            "Gemma3ForConditionalGeneration",
            "Gemma3nForCausalLM",
            "Gemma3nForConditionalGeneration",
        ]:
            # FIXME: https://github.com/sgl-project/sglang/pull/7367 is not compatible with gemma2 model.
            # It failed at this test: https://github.com/sgl-project/sglang/actions/runs/16255155597/job/45890331952#step:4:736
            logger.warning(
                f"Disable hybrid SWA memory for {model_arch} as it is not yet supported."
            )
            self.disable_hybrid_swa_memory = True

Lianmin Zheng's avatar
Lianmin Zheng committed
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
    def adjust_mem_fraction_for_vlm(self, model_config):
        vision_config = getattr(model_config.hf_config, "vision_config", None)
        if vision_config is None:
            return

        # roughly reduce the mem_fraction_static base on params of Vit
        original_server_arg_mem_fraction = self.mem_fraction_static
        # a base mem_fraction_static factor for regular Vit
        base_mem_fraction_reduction_ratio = 0.95

        vit_num_layers = getattr(vision_config, "num_hidden_layers", 24)
        vit_hidden_size = getattr(vision_config, "hidden_size", 1024)

        # baseline ViT params (ViT-L/14)
        baseline_vit_layers = 24
        baseline_vit_hidden_size = 1024

        # weight params count
        current_complexity_score = vit_num_layers * (vit_hidden_size**2)
        baseline_complexity_score = baseline_vit_layers * (baseline_vit_hidden_size**2)
        complexity_ratio = (
            current_complexity_score / baseline_complexity_score
            if baseline_complexity_score > 0
            else 1.0
        )

        # every time the complexity grows 100%, adjust final factor for 10%
        sensitivity_scale = 0.1
        dynamic_adjustment_factor = 1.0 - sensitivity_scale * (complexity_ratio - 1.0)
        dynamic_adjustment_factor = max(0.8, min(1.05, dynamic_adjustment_factor))

        final_overall_factor = (
            base_mem_fraction_reduction_ratio * dynamic_adjustment_factor
        )
        self.mem_fraction_static = (
            original_server_arg_mem_fraction * final_overall_factor
        )

Lianmin Zheng's avatar
Lianmin Zheng committed
2492

Lianmin Zheng's avatar
Lianmin Zheng committed
2493
def prepare_server_args(argv: List[str]) -> ServerArgs:
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
    """
    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
2506
    raw_args = parser.parse_args(argv)
2507
2508
2509
2510
    server_args = ServerArgs.from_cli_args(raw_args)
    return server_args


2511
2512
2513
ZMQ_TCP_PORT_DELTA = 233


Lianmin Zheng's avatar
Lianmin Zheng committed
2514
2515
@dataclasses.dataclass
class PortArgs:
2516
2517
2518
2519
2520
2521
    # 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
2522

2523
2524
    # The port for nccl initialization (torch.dist)
    nccl_port: int
2525

2526
2527
2528
    # The ipc filename for rpc call between Engine and Scheduler
    rpc_ipc_name: str

2529
2530
2531
    # The ipc filename for Scheduler to send metrics
    metrics_ipc_name: str

2532
2533
2534
    # The ipc filename for Tokenizer and worker tokenizer
    tokenizer_worker_ipc_name: Optional[str]

2535
    @staticmethod
2536
    def init_new(server_args, dp_rank: Optional[int] = None) -> "PortArgs":
2537
        if server_args.nccl_port is None:
Lianmin Zheng's avatar
Lianmin Zheng committed
2538
            nccl_port = server_args.port + random.randint(100, 1000)
2539
            while True:
Lianmin Zheng's avatar
Lianmin Zheng committed
2540
                if is_port_available(nccl_port):
2541
                    break
Lianmin Zheng's avatar
Lianmin Zheng committed
2542
2543
                if nccl_port < 60000:
                    nccl_port += 42
2544
                else:
Lianmin Zheng's avatar
Lianmin Zheng committed
2545
                    nccl_port -= 43
2546
        else:
Lianmin Zheng's avatar
Lianmin Zheng committed
2547
            nccl_port = server_args.nccl_port
2548

2549
2550
2551
2552
2553
2554
        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}",
Lianmin Zheng's avatar
Lianmin Zheng committed
2555
                nccl_port=nccl_port,
2556
                rpc_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
2557
                metrics_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
2558
                tokenizer_worker_ipc_name=None,
2559
2560
2561
2562
2563
            )
        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
2564
2565
2566
            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))
2567
2568
            else:
                dist_init_addr = server_args.dist_init_addr.split(":")
Vincent's avatar
Vincent committed
2569

2570
2571
2572
2573
2574
2575
            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
2576
2577
2578
            detokenizer_port = port_base + 1
            rpc_port = port_base + 2
            metrics_ipc_name = port_base + 3
2579
            if dp_rank is None:
2580
                # TokenizerManager to DataParallelController
2581
                scheduler_input_port = port_base + 4
2582
            else:
2583
                scheduler_input_port = port_base + 4 + 1 + dp_rank
2584
2585
2586
2587

            return PortArgs(
                tokenizer_ipc_name=f"tcp://{dist_init_host}:{port_base}",
                scheduler_input_ipc_name=f"tcp://{dist_init_host}:{scheduler_input_port}",
2588
                detokenizer_ipc_name=f"tcp://{dist_init_host}:{detokenizer_port}",
Lianmin Zheng's avatar
Lianmin Zheng committed
2589
                nccl_port=nccl_port,
2590
2591
                rpc_ipc_name=f"tcp://{dist_init_host}:{rpc_port}",
                metrics_ipc_name=f"tcp://{dist_init_host}:{metrics_ipc_name}",
2592
                tokenizer_worker_ipc_name=None,
2593
            )
2594

2595
2596
2597

class LoRAPathAction(argparse.Action):
    def __call__(self, parser, namespace, values, option_string=None):
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
        lora_paths = []
        if values:
            assert isinstance(values, list), "Expected a list of LoRA paths."
            for lora_path in values:
                lora_path = lora_path.strip()
                if lora_path.startswith("{") and lora_path.endswith("}"):
                    obj = json.loads(lora_path)
                    assert "lora_path" in obj and "lora_name" in obj, (
                        f"{repr(lora_path)} looks like a JSON str, "
                        "but it does not contain 'lora_name' and 'lora_path' keys."
                    )
                    lora_paths.append(obj)
                else:
                    lora_paths.append(lora_path)

        setattr(namespace, self.dest, lora_paths)
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623


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)
2624
2625


2626
2627
2628
2629
def print_deprecated_warning(message: str):
    logger.warning(f"\033[33m{message}\033[0m")


2630
def auto_choose_speculative_params(self: ServerArgs):
2631
2632
2633
2634
2635
    """
    Automatically choose the parameters for speculative decoding.

    You can tune them on your own models and prompts with scripts/playground/bench_speculative.py
    """
Lianmin Zheng's avatar
Lianmin Zheng committed
2636
    hf_config = self.get_hf_config()
2637
2638
    arch = hf_config.architectures[0]

2639
2640
2641
    if arch in ["LlamaForCausalLM"]:
        # The default value for llama
        return (5, 4, 8)
2642
2643
2644
2645
2646
2647
    elif arch in [
        "DeepseekV3ForCausalLM",
        "DeepseekV2ForCausalLM",
        "GptOssForCausalLM",
    ]:
        # The default value for deepseek and gpt-oss
2648
        return (3, 1, 4)
2649
2650
2651
2652
2653
    elif arch in ["Grok1ForCausalLM", "Grok1VForCausalLM"]:
        return (5, 4, 8)
    else:
        # The default value for all other models
        return (5, 4, 8)