server_args.py 132 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
Lianmin Zheng's avatar
Lianmin Zheng committed
14
15
"""The arguments of the server."""

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

25
from sglang.srt.connector import ConnectorType
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_npu,
40
    is_port_available,
41
    is_remote_url,
42
43
    is_sm90_supported,
    is_sm100_supported,
44
    is_triton_kernels_available,
45
    is_valid_ipv6_address,
46
    json_list_type,
bjmsong's avatar
bjmsong committed
47
    nullable_str,
48
    parse_connector_type,
49
)
50
from sglang.utils import is_in_ci
51

52
53
logger = logging.getLogger(__name__)

54
55
56
57
58
59
60
61
62
63
64
65
# Define constants
LOAD_FORMAT_CHOICES = [
    "auto",
    "pt",
    "safetensors",
    "npcache",
    "dummy",
    "sharded_state",
    "gguf",
    "bitsandbytes",
    "layered",
    "remote",
66
    "remote_instance",
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
]

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",
93
    "flex_attention",
94
95
96
    # NVIDIA specific
    "cutlass_mla",
    "fa3",
97
    "fa4",
98
99
100
101
102
103
104
105
106
107
108
109
110
    "flashinfer",
    "flashmla",
    "trtllm_mla",
    "trtllm_mha",
    "dual_chunk_flash_attn",
    # AMD specific
    "aiter",
    "wave",
    # Other platforms
    "intel_amx",
    "ascend",
]

111
112
LORA_BACKEND_CHOICES = ["triton", "csgmv"]

113
114
DISAGG_TRANSFER_BACKEND_CHOICES = ["mooncake", "nixl", "ascend", "fake"]

115
116
GRAMMAR_BACKEND_CHOICES = ["xgrammar", "outlines", "llguidance", "none"]

117
DETERMINISTIC_ATTENTION_BACKEND_CHOICES = ["flashinfer", "fa3", "triton"]
118

119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136

# 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)


137
138
139
140
def add_grammar_backend_choices(choices):
    GRAMMAR_BACKEND_CHOICES.extend(choices)


Lianmin Zheng's avatar
Lianmin Zheng committed
141
142
@dataclasses.dataclass
class ServerArgs:
Lianmin Zheng's avatar
Lianmin Zheng committed
143
    # Model and tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
144
145
146
    model_path: str
    tokenizer_path: Optional[str] = None
    tokenizer_mode: str = "auto"
147
    tokenizer_worker_num: int = 1
148
    skip_tokenizer_init: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
149
    load_format: str = "auto"
150
    model_loader_extra_config: str = "{}"
151
    trust_remote_code: bool = False
152
    context_length: Optional[int] = None
153
    is_embedding: bool = False
154
    enable_multimodal: Optional[bool] = None
155
    revision: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
156
    model_impl: str = "auto"
Lianmin Zheng's avatar
Lianmin Zheng committed
157

Lianmin Zheng's avatar
Lianmin Zheng committed
158
    # HTTP server
Lianmin Zheng's avatar
Lianmin Zheng committed
159
160
    host: str = "127.0.0.1"
    port: int = 30000
Lianmin Zheng's avatar
Lianmin Zheng committed
161
162
    skip_server_warmup: bool = False
    warmups: Optional[str] = None
163
    nccl_port: Optional[int] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
164

Lianmin Zheng's avatar
Lianmin Zheng committed
165
166
167
168
169
    # Quantization and data type
    dtype: str = "auto"
    quantization: Optional[str] = None
    quantization_param_path: Optional[str] = None
    kv_cache_dtype: str = "auto"
170
    enable_fp32_lm_head: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
171

Lianmin Zheng's avatar
Lianmin Zheng committed
172
    # Memory and scheduling
Lianmin Zheng's avatar
Lianmin Zheng committed
173
    mem_fraction_static: Optional[float] = None
174
    max_running_requests: Optional[int] = None
175
    max_queued_requests: Optional[int] = None
176
    max_total_tokens: Optional[int] = None
177
    chunked_prefill_size: Optional[int] = None
178
    max_prefill_tokens: int = 16384
179
    schedule_policy: str = "fcfs"
180
181
182
    enable_priority_scheduling: bool = False
    schedule_low_priority_values_first: bool = False
    priority_scheduling_preemption_threshold: int = 10
183
    schedule_conservativeness: float = 1.0
184
    page_size: Optional[int] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
185
186
187
    hybrid_kvcache_ratio: Optional[float] = None
    swa_full_tokens_ratio: float = 0.8
    disable_hybrid_swa_memory: bool = False
188
    radix_eviction_policy: str = "lru"
Lianmin Zheng's avatar
Lianmin Zheng committed
189

Lianmin Zheng's avatar
Lianmin Zheng committed
190
191
    # Runtime options
    device: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
192
    tp_size: int = 1
193
194
    pp_size: int = 1
    max_micro_batch_size: Optional[int] = None
195
    stream_interval: int = 1
196
    stream_output: bool = False
197
    random_seed: Optional[int] = None
198
    constrained_json_whitespace_pattern: Optional[str] = None
199
    watchdog_timeout: float = 300
200
    dist_timeout: Optional[int] = None  # timeout for torch.distributed
201
    download_dir: Optional[str] = None
202
    base_gpu_id: int = 0
203
    gpu_id_step: int = 1
204
    sleep_on_idle: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
205
206
207

    # Logging
    log_level: str = "info"
208
    log_level_http: Optional[str] = None
209
    log_requests: bool = False
210
    log_requests_level: int = 2
211
    crash_dump_folder: Optional[str] = None
Liangsheng Yin's avatar
Liangsheng Yin committed
212
    show_time_cost: bool = False
213
    enable_metrics: bool = False
214
    enable_metrics_for_all_schedulers: bool = False
215
216
    tokenizer_metrics_custom_labels_header: str = "x-custom-labels"
    tokenizer_metrics_allowed_custom_labels: Optional[List[str]] = None
217
218
    bucket_time_to_first_token: Optional[List[float]] = None
    bucket_inter_token_latency: Optional[List[float]] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
219
    bucket_e2e_request_latency: Optional[List[float]] = None
220
    collect_tokens_histogram: bool = False
221
222
    prompt_tokens_buckets: Optional[List[str]] = None
    generation_tokens_buckets: Optional[List[str]] = None
223
    decode_log_interval: int = 40
224
    enable_request_time_stats_logging: bool = False
225
    kv_events_config: Optional[str] = None
226
    gc_warning_threshold_secs: float = 0.0
227
228
    enable_trace: bool = False
    oltp_traces_endpoint: str = "localhost:4317"
Liangsheng Yin's avatar
Liangsheng Yin committed
229

230
    # API related
231
    api_key: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
232
    served_model_name: Optional[str] = None
233
    weight_version: str = "default"
Lianmin Zheng's avatar
Lianmin Zheng committed
234
235
    chat_template: Optional[str] = None
    completion_template: Optional[str] = None
236
    file_storage_path: str = "sglang_storage"
237
    enable_cache_report: bool = False
Xihuai Wang's avatar
Xihuai Wang committed
238
    reasoning_parser: Optional[str] = None
239
    tool_call_parser: Optional[str] = None
240
    tool_server: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
241

242
243
244
    # Data parallelism
    dp_size: int = 1
    load_balance_method: str = "round_robin"
245
    load_watch_interval: float = 0.1
246
247
    # FIXME: remove this after dp rank scheduling is fully supported with PD-Disaggregation
    prefill_round_robin_balance: bool = False
248

249
    # Multi-node distributed serving
250
    dist_init_addr: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
251
    nnodes: int = 1
252
    node_rank: int = 0
Lianmin Zheng's avatar
Lianmin Zheng committed
253
254
255

    # Model override args in JSON
    json_model_override_args: str = "{}"
256
    preferred_sampling_params: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
257

258
    # LoRA
259
    enable_lora: Optional[bool] = None
260
    max_lora_rank: Optional[int] = None
261
    lora_target_modules: Optional[Union[set[str], List[str]]] = None
262
263
264
    lora_paths: Optional[
        Union[dict[str, str], List[dict[str, str]], List[str], List[LoRARef]]
    ] = None
265
    max_loaded_loras: Optional[int] = None
266
    max_loras_per_batch: int = 8
267
    lora_backend: str = "triton"
268
    max_lora_chunk_size: Optional[int] = 16
269
270

    # Kernel backend
271
    attention_backend: Optional[str] = None
272
273
    decode_attention_backend: Optional[str] = None
    prefill_attention_backend: Optional[str] = None
274
    sampling_backend: Optional[str] = None
275
    grammar_backend: Optional[str] = None
276
    mm_attention_backend: Optional[str] = None
277

278
279
    # Speculative decoding
    speculative_algorithm: Optional[str] = None
280
    speculative_draft_model_path: Optional[str] = None
281
    speculative_draft_model_revision: Optional[str] = None
282
283
284
    speculative_num_steps: Optional[int] = None
    speculative_eagle_topk: Optional[int] = None
    speculative_num_draft_tokens: Optional[int] = None
285
286
    speculative_accept_threshold_single: float = 1.0
    speculative_accept_threshold_acc: float = 1.0
287
    speculative_token_map: Optional[str] = None
288
    speculative_attention_mode: str = "prefill"
289
290
291
292
293
294
295
296
    # For ngram only
    speculative_ngram_min_match_window_size: int = 1
    speculative_ngram_max_match_window_size: int = 12
    speculative_ngram_min_bfs_breadth: int = 1
    speculative_ngram_max_bfs_breadth: int = 10
    speculative_ngram_match_type: Literal["BFS", "PROB"] = "BFS"
    speculative_ngram_branch_length: int = 18
    speculative_ngram_capacity: int = 10 * 1000 * 1000
297

298
299
    # Expert parallelism
    ep_size: int = 1
300
301
302
303
304
305
306
307
308
    moe_a2a_backend: Literal["none", "deepep"] = "none"
    moe_runner_backend: Literal[
        "auto",
        "triton",
        "triton_kernel",
        "flashinfer_trtllm",
        "flashinfer_cutlass",
        "flashinfer_mxfp4",
    ] = "auto"
309
    flashinfer_mxfp4_moe_precision: Literal["default", "bf16"] = "default"
310
    enable_flashinfer_allreduce_fusion: bool = False
311
    deepep_mode: Literal["auto", "normal", "low_latency"] = "auto"
312
313
314
315
316
317
318
    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
319
    eplb_min_rebalancing_utilization_threshold: float = 1.0
320
321
322
323
324
325
326
327
    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
328
329
330
331
    # Mamba cache
    max_mamba_cache_size: Optional[int] = None
    mamba_ssm_dtype: str = "float32"

Lianmin Zheng's avatar
Lianmin Zheng committed
332
333
334
335
    # Hierarchical cache
    enable_hierarchical_cache: bool = False
    hicache_ratio: float = 2.0
    hicache_size: int = 0
336
    hicache_write_policy: str = "write_through"
337
338
    hicache_io_backend: str = "kernel"
    hicache_mem_layout: str = "layer_first"
Lianmin Zheng's avatar
Lianmin Zheng committed
339
    hicache_storage_backend: Optional[str] = None
pansicheng's avatar
pansicheng committed
340
    hicache_storage_prefetch_policy: str = "best_effort"
341
    hicache_storage_backend_extra_config: Optional[str] = None
342
343
    # LMCache
    enable_lmcache: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
344

345
346
    # Double Sparsity
    enable_double_sparsity: bool = False
Vincent's avatar
Vincent committed
347
    ds_channel_config_path: Optional[str] = None
348
349
350
351
352
    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
353
354
355
356
357
358
359
    # 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"

360
    # Optimization/debug options
Lianmin Zheng's avatar
Lianmin Zheng committed
361
    disable_radix_cache: bool = False
362
363
    cuda_graph_max_bs: Optional[int] = None
    cuda_graph_bs: Optional[List[int]] = None
364
    disable_cuda_graph: bool = False
365
    disable_cuda_graph_padding: bool = False
366
    enable_profile_cuda_graph: bool = False
367
    enable_cudagraph_gc: bool = False
368
    enable_nccl_nvls: bool = False
369
    enable_symm_mem: bool = False
370
    disable_flashinfer_cutlass_moe_fp4_allgather: bool = False
371
    enable_tokenizer_batch_encode: bool = False
372
    disable_outlines_disk_cache: bool = False
373
    disable_custom_all_reduce: bool = False
374
    enable_mscclpp: bool = False
375
    disable_overlap_schedule: bool = False
376
    enable_mixed_chunk: bool = False
Ke Bao's avatar
Ke Bao committed
377
    enable_dp_attention: bool = False
378
    enable_dp_lm_head: bool = False
379
    enable_two_batch_overlap: bool = False
380
    enable_single_batch_overlap: bool = False
381
    tbo_token_distribution_threshold: float = 0.48
382
    enable_torch_compile: bool = False
383
    torch_compile_max_bs: int = 32
384
    torchao_config: str = ""
385
    enable_nan_detection: bool = False
386
    enable_p2p_check: bool = False
387
    triton_attention_reduce_in_fp32: bool = False
388
    triton_attention_num_kv_splits: int = 8
389
    triton_attention_split_tile_size: Optional[int] = None
390
    num_continuous_decode_steps: int = 1
391
    delete_ckpt_after_loading: bool = False
392
    enable_memory_saver: bool = False
393
    allow_auto_truncate: bool = False
394
    enable_custom_logit_processor: bool = False
395
    flashinfer_mla_disable_ragged: bool = False
396
    disable_shared_experts_fusion: bool = False
397
    disable_chunked_prefix_cache: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
398
    disable_fast_image_processor: bool = False
399
    keep_mm_feature_on_device: bool = False
400
    enable_return_hidden_states: bool = False
401
    scheduler_recv_interval: int = 1
402
    numa_node: Optional[List[int]] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
403
    enable_deterministic_inference: bool = False
404

405
406
407
408
409
    # Dynamic batch tokenizer
    enable_dynamic_batch_tokenizer: bool = False
    dynamic_batch_tokenizer_batch_size: int = 32
    dynamic_batch_tokenizer_batch_timeout: float = 0.002

410
411
412
413
    # 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
414
    debug_tensor_dump_prefill_only: bool = False
415

Lianmin Zheng's avatar
Lianmin Zheng committed
416
    # PD disaggregation: can be "null" (not disaggregated), "prefill" (prefill-only), or "decode" (decode-only)
417
    disaggregation_mode: Literal["null", "prefill", "decode"] = "null"
418
    disaggregation_transfer_backend: str = "mooncake"
419
    disaggregation_bootstrap_port: int = 8998
Byron Hsu's avatar
Byron Hsu committed
420
421
422
    disaggregation_decode_tp: Optional[int] = None
    disaggregation_decode_dp: Optional[int] = None
    disaggregation_prefill_pp: Optional[int] = 1
423
    disaggregation_ib_device: Optional[str] = None
424
    disaggregation_decode_enable_offload_kvcache: bool = False
425
    num_reserved_decode_tokens: int = 512  # used for decode kv cache offload in PD
426
427
428
    # FIXME: hack to reduce ITL when decode bs is small
    disaggregation_decode_polling_interval: int = 1

Lianmin Zheng's avatar
Lianmin Zheng committed
429
    # For model weight update and weight loading
430
    custom_weight_loader: Optional[List[str]] = None
431
    weight_loader_disable_mmap: bool = False
432
433
434
435
    remote_instance_weight_loader_seed_instance_ip: Optional[str] = None
    remote_instance_weight_loader_seed_instance_service_port: Optional[int] = None
    remote_instance_weight_loader_send_weights_group_ports: Optional[List[int]] = None

436
437
438
439
    # For PD-Multiplexing
    enable_pdmux: bool = False
    sm_group_num: int = 3

Lianmin Zheng's avatar
Lianmin Zheng committed
440
441
442
443
444
445
    def __post_init__(self):
        """
        Orchestrates the handling of various server arguments, ensuring proper configuration and validation.
        """
        # Handle deprecated arguments.
        self._handle_deprecated_args()
Yi Zhang's avatar
Yi Zhang committed
446

Lianmin Zheng's avatar
Lianmin Zheng committed
447
448
449
450
451
452
        # Set missing default values.
        self._handle_missing_default_values()

        # Get GPU memory capacity, which is a common dependency for several configuration steps.
        gpu_mem = get_device_memory_capacity(self.device)

453
454
        # Handle memory-related, chunked prefill, and CUDA graph batch size configurations.
        self._handle_gpu_memory_settings(gpu_mem)
Lianmin Zheng's avatar
Lianmin Zheng committed
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504

        # Handle device-specific backends.
        self._handle_hpu_backends()
        self._handle_cpu_backends()

        # Apply model-specific adjustments.
        self._handle_model_specific_adjustments()

        # Set kernel backends.
        self._handle_sampling_backend()
        self._handle_attention_backend_compatibility()
        self._handle_page_size()
        self._handle_amd_specifics()
        self._handle_grammar_backend()

        # Handle data parallelism.
        self._handle_data_parallelism()

        # Handle MoE configurations.
        self._handle_moe_kernel_config()
        self._handle_deepep_moe()
        self._handle_eplb_and_dispatch()
        self._handle_expert_distribution_metrics()

        # Handle pipeline parallelism.
        self._handle_pipeline_parallelism()

        # Handle Hicache settings.
        self._handle_hicache()

        # Handle speculative decoding logic.
        self._handle_speculative_decoding()

        # Handle model loading format.
        self._handle_load_format()

        # Handle PD disaggregation.
        self._handle_disaggregation()

        # Validate tokenizer settings.
        self._handle_tokenizer_batching()

        # Propagate environment variables.
        self._handle_environment_variables()

        # Validate cache settings.
        self._handle_cache_compatibility()

        # Validate metrics labels.
        self._handle_metrics_labels()
505

Lianmin Zheng's avatar
Lianmin Zheng committed
506
507
508
509
510
        # Handle deterministic inference.
        self._handle_deterministic_inference()

        # Handle any other necessary validations.
        self._handle_other_validations()
511

512
    def _handle_deprecated_args(self):
Lianmin Zheng's avatar
Lianmin Zheng committed
513
        pass
514

515
    def _handle_missing_default_values(self):
Lianmin Zheng's avatar
Lianmin Zheng committed
516
517
        if self.tokenizer_path is None:
            self.tokenizer_path = self.model_path
518
519
        if self.served_model_name is None:
            self.served_model_name = self.model_path
520
521
        if self.device is None:
            self.device = get_device()
522
523
524
        if self.random_seed is None:
            self.random_seed = random.randint(0, 1 << 30)

525
526
    def _handle_gpu_memory_settings(self, gpu_mem):
        """
527
528
        Configure GPU memory-dependent settings including
        chunked_prefill_size, cuda_graph_max_bs, and mem_fraction_static.
529

530
531
532
533
534
        Here are our heuristics:
        - Set chunked_prefill_size and cuda_graph_max_bs based on the GPU memory capacity.
          This is because GPUs with more memory are generally more powerful, we need to use a larger
          chunked_prefill_size and a larger cuda_graph_max_bs to fully utilize the GPU.
        - Then set mem_fraction_static based on chunked_prefill_size and cuda_graph_max_bs.
535

536
          GPU memory capacity = model weights + KV cache pool + activations + cuda graph buffers
537

538
539
          The argument mem_fraction_static is defined as (model weights + KV cache pool) / GPU memory capacity,
          or equivalently, mem_fraction_static = (GPU memory capacity - activations - cuda graph buffers) / GPU memory capacity.
Lianmin Zheng's avatar
Lianmin Zheng committed
540

541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
          In order to compute mem_fraction_static, we need to estimate the size of activations and cuda graph buffers.
          The activation memory is proportional to the chunked_prefill_size.
          The cuda graph memory is proportional to the cuda_graph_max_bs.
          We use reserved_mem = chunked_prefill_size * 1.5 + cuda_graph_max_bs * 2 to estimate the size of activations and cuda graph buffers in GB.
          and set mem_fraction_static = (GPU memory capacity - reserved_mem) / GPU memory capacity.

          The coefficient 1.5 is a heuristic value, in the future, we can do better estimation by looking at the model types, hidden sizes or even do a dummy run.
        """
        if gpu_mem is not None:
            if gpu_mem < 20 * 1024:
                # T4, 4080
                # (chunked_prefill_size 2k, cuda_graph_max_bs 8)
                if self.chunked_prefill_size is None:
                    self.chunked_prefill_size = 2048
                if self.cuda_graph_max_bs is None:
556
                    self.cuda_graph_max_bs = 8
557
558
559
560
561
562
563
            elif gpu_mem < 35 * 1024:
                # A10, 4090, 5090
                # (chunked_prefill_size 2k, cuda_graph_max_bs 16 if tp < 4 else 80)
                if self.chunked_prefill_size is None:
                    self.chunked_prefill_size = 2048
                if self.cuda_graph_max_bs is None:
                    # Based on detailed statistics, when serving TP1/TP2 models on lower-end GPUs with HBM < 35GB, 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.
564
565
566
567
568
569
                    # 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 self.tp_size < 4:
                        self.cuda_graph_max_bs = 16
                    else:
                        self.cuda_graph_max_bs = 80
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
            elif gpu_mem < 60 * 1024:
                # A100 (40GB), L40,
                # (chunked_prefill_size 4k, cuda_graph_max_bs 32 if tp < 4 else 160)
                if self.chunked_prefill_size is None:
                    self.chunked_prefill_size = 4096
                if self.cuda_graph_max_bs is None:
                    if self.tp_size < 4:
                        self.cuda_graph_max_bs = 32
                    else:
                        self.cuda_graph_max_bs = 160
            elif gpu_mem < 90 * 1024:
                # H100, A100
                # (chunked_prefill_size 8k, cuda_graph_max_bs 256 if tp < 4 else 512)
                if self.chunked_prefill_size is None:
                    self.chunked_prefill_size = 8192
                if self.cuda_graph_max_bs is None:
                    if self.tp_size < 4:
                        self.cuda_graph_max_bs = 256
                    else:
                        self.cuda_graph_max_bs = 512
            elif gpu_mem < 160 * 1024:
                # H20, H200
                # (chunked_prefill_size 8k, cuda_graph_max_bs 256 if tp < 4 else 512)
                if self.chunked_prefill_size is None:
                    self.chunked_prefill_size = 8192
                if self.cuda_graph_max_bs is None:
596
597
598
599
600
                    if self.tp_size < 4:
                        self.cuda_graph_max_bs = 256
                    else:
                        self.cuda_graph_max_bs = 512
            else:
601
602
603
604
605
606
607
608
609
610
611
                # B200, MI300
                # (chunked_prefill_size 16k, cuda_graph_max_bs 512)
                if self.chunked_prefill_size is None:
                    self.chunked_prefill_size = 16384
                if self.cuda_graph_max_bs is None:
                    self.cuda_graph_max_bs = 512
        else:
            # Fallback defaults when gpu_mem is None
            if self.chunked_prefill_size is None:
                self.chunked_prefill_size = 4096
            if self.cuda_graph_max_bs is None:
612
613
                self.cuda_graph_max_bs = 160

614
        # Set cuda graph batch sizes
615
616
        if self.cuda_graph_bs is None:
            self.cuda_graph_bs = self._generate_cuda_graph_batch_sizes()
617
618
619
620
621
        else:
            self.cuda_graph_max_bs = max(self.cuda_graph_bs)

        if self.mem_fraction_static is None:
            # Constant meta data (e.g., from attention backend)
622
            reserved_mem = 512
623
624
625
626
627
628
629
630
            # For activation during large prefill
            if self.chunked_prefill_size > 0:
                reserved_mem += max(self.chunked_prefill_size, 2048) * 1.5
            else:
                reserved_mem += max(self.max_prefill_tokens, 2048) * 1.5
            # For cuda graphs
            reserved_mem += self.cuda_graph_max_bs * 2
            # Some adjustments for large parallel size
631
            reserved_mem += self.tp_size * self.pp_size / 8 * 1024
632
633
634
635
636
637
638
639
640
641
642

            if self.enable_dp_attention:
                # DP attention needs more padding for some operations
                reserved_mem += self.cuda_graph_max_bs * self.dp_size * 3

                # DP attention uses much more memory for large cuda graph max bs,
                # likely due to some inefficiencies in torch allocator or our implementation.
                # So we need to reserve more memory.
                if self.cuda_graph_max_bs > 300:
                    reserved_mem += self.cuda_graph_max_bs * self.dp_size * 1.5

643
            if gpu_mem is not None and gpu_mem > 60 * 1024:
644
645
646
647
648
649
650
651
652
653
                reserved_mem = max(reserved_mem, 10 * 1024)

            if self.speculative_algorithm is not None:
                if self.speculative_algorithm == "STANDALONE":
                    # standalonedraft model and cuda graphs
                    reserved_mem += 6 * 1024
                elif self.speculative_algorithm != "NGRAM":
                    # eagle draft models and cuda graphs
                    reserved_mem += 2 * 1024

654
655
656
657
658
            self.mem_fraction_static = (
                round((gpu_mem - reserved_mem) / gpu_mem, 3)
                if gpu_mem is not None
                else 0.88
            )
659
660
661
662
663
664
665
666

            # Lazy init to avoid circular import
            # Multimodal models need more memory for the image processor
            from sglang.srt.configs.model_config import ModelConfig

            model_config = ModelConfig.from_server_args(self)
            if model_config.is_multimodal:
                self.adjust_mem_fraction_for_vlm(model_config)
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681

    def _generate_cuda_graph_batch_sizes(self):
        """
        Generate the list of batch sizes for CUDA graph capture based on cuda_graph_max_bs.
        This integrates the logic from cuda_graph_runner.py.
        """
        # Handle disable_cuda_graph_padding as the first condition for both spec and non-spec
        if self.disable_cuda_graph_padding:
            capture_bs = list(range(1, self.cuda_graph_max_bs + 1))
        elif self.speculative_algorithm is None:
            # Normal case: [1, 2, 4, 8, 12] + list(range(16, 257, 8)) + list(range(272, 512, 16)) + list(range(512, cuda_graph_max_bs + 1))
            capture_bs = (
                [1, 2, 4, 8, 12]
                + list(range(16, 257, 8))
                + list(range(272, 512, 16))
682
                + list(range(512, self.cuda_graph_max_bs + 1, 32))
683
684
685
686
687
688
689
690
691
692
693
694
695
696
            )
        else:
            # Spec decoding case: list(range(1, 9, 1)) + list(range(10, 33, 2)) + list(range(40, 64, 4)) + list(range(72, 257, 8))
            capture_bs = (
                list(range(1, 9, 1))
                + list(range(10, 33, 2))
                + list(range(40, 64, 4))
                + list(range(72, 257, 8))
                + list(range(272, self.cuda_graph_max_bs + 1, 16))
            )

        capture_bs = [bs for bs in capture_bs if bs <= self.cuda_graph_max_bs]

        return capture_bs
697

698
    def _handle_hpu_backends(self):
699
700
701
702
        if self.device == "hpu":
            self.attention_backend = "torch_native"
            self.sampling_backend = "pytorch"

703
    def _handle_cpu_backends(self):
704
705
706
707
708
        if self.device == "cpu":
            if self.attention_backend is None:
                self.attention_backend = "intel_amx"
            self.sampling_backend = "pytorch"

Lianmin Zheng's avatar
Lianmin Zheng committed
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
    def _handle_model_specific_adjustments(self):
        if parse_connector_type(self.model_path) == ConnectorType.INSTANCE:
            return

        hf_config = self.get_hf_config()
        model_arch = hf_config.architectures[0]
        if model_arch in ["GptOssForCausalLM"]:
            if self.attention_backend is None:
                if is_cuda() and is_sm100_supported():
                    self.attention_backend = "trtllm_mha"
                elif is_cuda() and is_sm90_supported():
                    self.attention_backend = "fa3"
                else:
                    self.attention_backend = "triton"
            supported_backends = ["triton", "trtllm_mha", "fa3"]
            logger.info(
                f"Use {self.attention_backend} as attention backend for GptOssForCausalLM"
            )
            assert (
                self.attention_backend in supported_backends
            ), f"GptOssForCausalLM requires one of {supported_backends} attention backend, but got '{self.attention_backend}'"

            if is_sm100_supported():
                if not self.enable_dp_attention:
                    self.enable_flashinfer_allreduce_fusion = True
                    logger.info(
                        "Enable FlashInfer AllReduce Fusion on sm100 for GptOssForCausalLM"
                    )
            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:
                self.moe_runner_backend = "flashinfer_mxfp4"
                logger.warning(
                    "Detected SM100 and MXFP4 quantization format for GPT-OSS model, enabling FlashInfer MXFP4 MOE kernel."
                )
            else:
                if self.moe_runner_backend == "triton_kernel":
                    assert (
                        self.ep_size == 1
                    ), "Triton kernel MoE is only supported when ep_size == 1"
                if (
                    self.moe_runner_backend == "auto"
                    and self.ep_size == 1
                    and is_triton_kernels_available()
                ):
                    self.moe_runner_backend = "triton_kernel"
                    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"

        elif "Llama4" in model_arch and self.device != "cpu":
            assert self.attention_backend in {
                "fa3",
                "aiter",
                "triton",
            }, "fa3, aiter, or triton is required for Llama4 model"
        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

787
    def _handle_sampling_backend(self):
788
        if self.sampling_backend is None:
789
790
791
792
            self.sampling_backend = (
                "flashinfer" if is_flashinfer_available() else "pytorch"
            )

793
    def _handle_attention_backend_compatibility(self):
794
        if self.attention_backend == "torch_native":
795
            logger.warning(
796
797
798
                "Cuda graph is disabled because of using torch native attention backend"
            )
            self.disable_cuda_graph = True
799

800
801
802
803
804
805
806
807
808
        if self.attention_backend == "flex_attention":
            logger.warning(
                "Cuda graph is disabled because of using torch Flex Attention backend"
            )
            self.disable_cuda_graph = True
            assert (
                self.speculative_algorithm is None
            ), "Speculative decoding is currently not supported with Flex Attention backend"

809
        if is_npu() and self.attention_backend in ["ascend"]:
810
811
812
813
814
            logger.warning(
                "At this moment Ascend attention backend only supports a page_size of 128, change page_size to 128."
            )
            self.page_size = 128

815
816
817
818
        if (
            self.attention_backend == "flashmla"
            or self.decode_attention_backend == "flashmla"
        ):
Lianmin Zheng's avatar
Lianmin Zheng committed
819
820
821
822
823
            logger.warning(
                "FlashMLA only supports a page_size of 64, change page_size to 64."
            )
            self.page_size = 64

824
825
826
827
        if (
            self.attention_backend == "cutlass_mla"
            or self.decode_attention_backend == "cutlass_mla"
        ):
Lianmin Zheng's avatar
Lianmin Zheng committed
828
829
830
831
832
            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
833
834
835
836
        if (
            self.attention_backend == "trtllm_mla"
            or self.decode_attention_backend == "trtllm_mla"
        ):
837
838
839
840
841
842
843
844
845
846
            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
847
848
849
850
851

            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."
                )
852

853
854
855
856
857
        if (
            self.attention_backend == "trtllm_mha"
            or self.decode_attention_backend == "trtllm_mha"
            or self.prefill_attention_backend == "trtllm_mha"
        ):
858
859
860
861
862
863
864
865
866
867
868
            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

869
870
        if self.attention_backend == "dual_chunk_flash_attn":
            logger.warning(
871
                "Mixed chunk, radix cache, and cuda graphs are disabled because of using dual chunk flash attention backend"
872
873
874
875
876
            )
            self.enable_mixed_chunk = False
            self.disable_cuda_graph = True
            self.disable_radix_cache = True

877
    def _handle_page_size(self):
878
879
880
        if self.page_size is None:
            self.page_size = 1

881
    def _handle_amd_specifics(self):
882
883
884
        if is_hip():
            self.triton_attention_num_kv_splits = 16

885
    def _handle_grammar_backend(self):
886
887
        if self.grammar_backend is None:
            self.grammar_backend = "xgrammar"
888

889
    def _handle_data_parallelism(self):
890
891
        if self.dp_size == 1:
            self.enable_dp_attention = False
892
            self.enable_dp_lm_head = False
893

Ke Bao's avatar
Ke Bao committed
894
        if self.enable_dp_attention:
895
            self.schedule_conservativeness = self.schedule_conservativeness * 0.3
Lianmin Zheng's avatar
Lianmin Zheng committed
896
897
            assert self.tp_size % self.dp_size == 0
            self.chunked_prefill_size = self.chunked_prefill_size // self.dp_size
898
            logger.warning(
899
                f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
900
            )
901

902
903
904
        if self.enable_dp_lm_head:
            assert (
                self.enable_dp_attention
905
            ), "Please enable dp attention when setting enable_dp_lm_head. "
906

907
    def _handle_moe_kernel_config(self):
908
        if self.moe_runner_backend == "flashinfer_cutlass":
909
910
911
            assert (
                self.quantization == "modelopt_fp4"
            ), "modelopt_fp4 quantization is required for Flashinfer MOE"
912
913
914
915
            assert self.ep_size in [
                1,
                self.tp_size,
            ], "The expert parallel size must be 1 or the same as the tensor parallel size"
916

917
        if self.moe_runner_backend == "flashinfer_trtllm":
918
919
            assert (
                self.quantization == "modelopt_fp4" or self.quantization == "fp8"
920
            ), "modelopt_fp4 or fp8 quantization is required for Flashinfer TRTLLM MoE"
921
922
923
924
            self.disable_shared_experts_fusion = True
            logger.warning(
                "FlashInfer TRTLLM MoE is enabled. --disable-shared-experts-fusion is automatically set."
            )
925

926
    def _handle_deepep_moe(self):
927
        if self.moe_a2a_backend == "deepep":
928
929
930
            if self.deepep_mode == "normal":
                logger.warning("Cuda graph is disabled because deepep_mode=`normal`")
                self.disable_cuda_graph = True
931
            self.ep_size = self.tp_size
Lianmin Zheng's avatar
Lianmin Zheng committed
932
            logger.warning(
933
934
                f"DeepEP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
            )
935

936
    def _handle_eplb_and_dispatch(self):
937
938
        if self.enable_eplb and (self.expert_distribution_recorder_mode is None):
            self.expert_distribution_recorder_mode = "stat"
939
            logger.warning(
940
                "EPLB is enabled. The expert_distribution_recorder_mode is automatically set."
941
942
943
944
945
946
947
            )

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

948
        if self.enable_eplb:
949
            assert self.ep_size > 1
950

951
    def _handle_expert_distribution_metrics(self):
952
953
954
955
956
        if self.enable_expert_distribution_metrics and (
            self.expert_distribution_recorder_mode is None
        ):
            self.expert_distribution_recorder_mode = "stat"

957
        if self.expert_distribution_recorder_buffer_size is None:
958
959
            if (x := self.eplb_rebalance_num_iterations) is not None:
                self.expert_distribution_recorder_buffer_size = x
960
961
962
            elif self.expert_distribution_recorder_mode is not None:
                self.expert_distribution_recorder_buffer_size = 1000

963
    def _handle_pipeline_parallelism(self):
Lianmin Zheng's avatar
Lianmin Zheng committed
964
965
966
967
968
969
        if self.pp_size > 1:
            self.disable_overlap_schedule = True
            logger.warning(
                "Pipeline parallelism is incompatible with overlap schedule."
            )

970
    def _handle_hicache(self):
971
        if self.hicache_storage_backend == "mooncake":
972
973
974
975
976
977
978
979
980
            if self.hicache_mem_layout == "layer_first":
                if self.hicache_io_backend == "direct":
                    self.hicache_mem_layout = "page_first_direct"
                elif self.hicache_io_backend == "kernel":
                    self.hicache_mem_layout = "page_first"
                logger.warning(
                    f"Mooncake storage backend does not support layer_first layout, "
                    f"switching to {self.hicache_mem_layout} layout for {self.hicache_io_backend} io backend"
                )
981

982
983
984
985
986
987
988
        if self.hicache_mem_layout == "page_first_direct":
            if self.hicache_io_backend != "direct":
                self.hicache_io_backend = "direct"
                logger.warning(
                    "Page first direct layout only support direct io backend"
                )

989
    def _handle_speculative_decoding(self):
990
991
992
        if self.speculative_algorithm == "NEXTN":
            self.speculative_algorithm = "EAGLE"

993
        if self.speculative_algorithm in ("EAGLE", "EAGLE3", "STANDALONE"):
994
            if self.speculative_algorithm == "STANDALONE" and self.enable_dp_attention:
995
                # TODO: support dp attention for standalone speculative decoding
996
997
998
                raise ValueError(
                    "Currently standalone speculative decoding does not support dp attention."
                )
999
            if self.max_running_requests is None:
1000
                self.max_running_requests = 48
1001
            self.disable_overlap_schedule = True
Lianmin Zheng's avatar
Lianmin Zheng committed
1002
            logger.warning(
1003
                "Overlap scheduler is disabled because of using "
1004
                "eagle speculative decoding."
1005
            )
1006
1007
1008
1009
1010
1011
            if self.enable_mixed_chunk:
                self.enable_mixed_chunk = False
                logger.warning(
                    "Mixed chunked prefill is disabled because of using "
                    "eagle speculative decoding."
                )
1012

Lianmin Zheng's avatar
Lianmin Zheng committed
1013
            model_arch = self.get_hf_config().architectures[0]
strgrb's avatar
strgrb committed
1014
1015
1016
            if model_arch in [
                "DeepseekV3ForCausalLM",
                "Glm4MoeForCausalLM",
Yuan Luo's avatar
Yuan Luo committed
1017
                "BailingMoeForCausalLM",
strgrb's avatar
strgrb committed
1018
1019
                "BailingMoeV2ForCausalLM",
            ]:
1020
1021
1022
1023
1024
1025
                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."
                    )
1026

1027
1028
1029
1030
1031
1032
1033
1034
1035
            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,
1036
                ) = auto_choose_speculative_params(self)
1037

1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
            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."
                    )

1048
1049
1050
1051
            if (
                self.speculative_eagle_topk == 1
                and self.speculative_num_draft_tokens != self.speculative_num_steps + 1
            ):
Lianmin Zheng's avatar
Lianmin Zheng committed
1052
                logger.warning(
1053
1054
1055
                    "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
1056

1057
1058
1059
1060
1061
1062
1063
1064
1065
            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."
                )

1066
        if self.speculative_algorithm == "NGRAM":
1067
1068
            if not self.device.startswith("cuda"):
                raise ValueError(
1069
                    "Ngram speculative decoding only supports CUDA device."
1070
1071
1072
1073
1074
                )
            if self.max_running_requests is None:
                self.max_running_requests = 48
            self.disable_overlap_schedule = True
            self.enable_mixed_chunk = False
1075
            self.speculative_eagle_topk = self.speculative_ngram_max_bfs_breadth
1076
1077
            if self.speculative_num_draft_tokens is None:
                self.speculative_num_draft_tokens = (
1078
                    self.speculative_ngram_max_match_window_size
1079
1080
1081
                )
            logger.warning(
                "The overlap scheduler and mixed chunked prefill are disabled because of "
1082
                "using ngram speculative decoding."
1083
            )
1084

1085
1086
1087
1088
1089
1090
            if (
                self.speculative_eagle_topk > 1
                and self.page_size > 1
                and self.attention_backend != "flashinfer"
            ):
                raise ValueError(
1091
1092
1093
1094
                    f"speculative_eagle_topk({self.speculative_eagle_topk}) > 1 "
                    f"with page_size({self.page_size}) > 1 is unstable "
                    "and produces incorrect results for paged attention backends. "
                    "This combination is only supported for the 'flashinfer' backend."
1095
1096
                )
            if self.enable_dp_attention:
1097
                # TODO: support dp attention for ngram speculative decoding
1098
                raise ValueError(
1099
                    "Currently ngram speculative decoding does not support dp attention."
1100
                )
1101
1102

    def _handle_load_format(self):
1103
1104
1105
1106
1107
        if (
            self.load_format == "auto" or self.load_format == "gguf"
        ) and check_gguf_file(self.model_path):
            self.quantization = self.load_format = "gguf"

1108
1109
        if is_remote_url(self.model_path):
            self.load_format = "remote"
1110

1111
1112
        if self.custom_weight_loader is None:
            self.custom_weight_loader = []
1113

1114
1115
1116
1117
1118
1119
1120
1121
        if self.load_format == "remote_instance":
            if (
                self.remote_instance_weight_loader_seed_instance_ip is None
                or self.remote_instance_weight_loader_seed_instance_service_port is None
                or self.remote_instance_weight_loader_send_weights_group_ports is None
            ):
                self.load_format = "auto"

1122
    def _handle_disaggregation(self):
Byron Hsu's avatar
Byron Hsu committed
1123
1124
1125
1126
1127
1128
1129
1130
        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
1131
            self.disable_radix_cache = True
1132
            logger.warning("KV cache is forced as chunk cache for decode server")
1133
1134
1135
1136
1137
1138
1139

            if self.dp_size > 1 and not is_in_ci():
                assert self.prefill_round_robin_balance, (
                    "Prefill round robin balance is required when dp size > 1. "
                    "Please make sure that the prefill instance is launched with `--load-balance-method round_robin`"
                    " and `--prefill-round-robin-balance` is set for decode server."
                )
Byron Hsu's avatar
Byron Hsu committed
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
        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
1150

1151
    def _handle_tokenizer_batching(self):
1152
1153
1154
1155
1156
1157
        if self.enable_tokenizer_batch_encode and self.enable_dynamic_batch_tokenizer:
            raise ValueError(
                "Cannot enable both --enable-tokenizer-batch-encode and --enable-dynamic-batch-tokenizer. "
                "Please choose one tokenizer batching approach."
            )

1158
    def _handle_environment_variables(self):
1159
1160
1161
        os.environ["SGLANG_ENABLE_TORCH_COMPILE"] = (
            "1" if self.enable_torch_compile else "0"
        )
Yi Zhang's avatar
Yi Zhang committed
1162
        os.environ["SGLANG_MAMBA_SSM_DTYPE"] = self.mamba_ssm_dtype
1163
1164
1165
        os.environ["SGLANG_DISABLE_OUTLINES_DISK_CACHE"] = (
            "1" if self.disable_outlines_disk_cache else "0"
        )
1166
1167
1168
        os.environ["SGLANG_ENABLE_DETERMINISTIC_INFERENCE"] = (
            "1" if self.enable_deterministic_inference else "0"
        )
1169

1170
    def _handle_cache_compatibility(self):
1171
1172
1173
1174
1175
1176
        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."
            )

1177
1178
1179
1180
1181
1182
1183
1184
        if (
            self.disaggregation_decode_enable_offload_kvcache
            and self.disaggregation_mode != "decode"
        ):
            raise ValueError(
                "The argument disaggregation-decode-enable-offload-kvcache is only supported for decode side."
            )

1185
    def _handle_metrics_labels(self):
1186
1187
        if (
            not self.tokenizer_metrics_custom_labels_header
1188
            and self.tokenizer_metrics_allowed_custom_labels
1189
1190
        ):
            raise ValueError(
1191
                "Please set --tokenizer-metrics-custom-labels-header when setting --tokenizer-metrics-allowed-custom-labels."
1192
1193
            )

1194
    def _handle_deterministic_inference(self):
1195
        if self.enable_deterministic_inference:
1196
            # Check sampling backend
1197
1198
1199
1200
            self.sampling_backend = "pytorch"
            logger.warning(
                "Sampling backend is set to pytorch for deterministic inference."
            )
1201
1202
1203
1204
1205
1206
1207

            # Check attention backend
            if self.attention_backend not in DETERMINISTIC_ATTENTION_BACKEND_CHOICES:
                raise ValueError(
                    f"Currently only {DETERMINISTIC_ATTENTION_BACKEND_CHOICES} attention backends are supported for deterministic inference."
                )

1208
            # Currently, only FA3 supports radix cache. Support for other backends is in progress
1209
1210
1211
            if self.attention_backend != "fa3":
                self.disable_radix_cache = True
                logger.warning(
1212
                    f"Currently radix cache is not compatible with {self.attention_backend} attention backend for deterministic inference. It will be supported in the future."
1213
                )
1214
1215
1216

            # Check TP size
            if self.tp_size > 1:
1217
1218
1219
1220
                os.environ["NCCL_ALGO"] = "allreduce:tree"
                self.disable_custom_all_reduce = True
                logger.warning(
                    "NCCL_ALGO is set to 'allreduce:tree' and custom all reduce is disabled for deterministic inference when TP size > 1."
1221
1222
                )

1223
    def _handle_other_validations(self):
fzyzcjy's avatar
fzyzcjy committed
1224
        pass
1225

Lianmin Zheng's avatar
Lianmin Zheng committed
1226
1227
    @staticmethod
    def add_cli_args(parser: argparse.ArgumentParser):
Lianmin Zheng's avatar
Lianmin Zheng committed
1228
        # Model and tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
1229
1230
        parser.add_argument(
            "--model-path",
1231
            "--model",
Lianmin Zheng's avatar
Lianmin Zheng committed
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
            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.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1242
1243
1244
1245
1246
1247
1248
1249
1250
        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.",
        )
1251
1252
1253
1254
1255
1256
        parser.add_argument(
            "--tokenizer-worker-num",
            type=int,
            default=ServerArgs.tokenizer_worker_num,
            help="The worker num of the tokenizer manager.",
        )
1257
1258
1259
        parser.add_argument(
            "--skip-tokenizer-init",
            action="store_true",
1260
            help="If set, skip init tokenizer and pass input_ids in generate request.",
1261
        )
1262
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
1263
1264
1265
            "--load-format",
            type=str,
            default=ServerArgs.load_format,
1266
            choices=LOAD_FORMAT_CHOICES,
Lianmin Zheng's avatar
Lianmin Zheng committed
1267
1268
1269
1270
1271
1272
1273
1274
1275
            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, '
1276
            "which is mainly for profiling."
1277
1278
            '"gguf" will load the weights in the gguf format. '
            '"bitsandbytes" will load the weights using bitsandbytes '
1279
1280
1281
1282
            "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
1283
        )
1284
1285
1286
1287
1288
1289
1290
        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,
        )
1291
1292
1293
1294
1295
        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
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
        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
1367
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
1368
            "--dtype",
Cody Yu's avatar
Cody Yu committed
1369
            type=str,
Lianmin Zheng's avatar
Lianmin Zheng committed
1370
            default=ServerArgs.dtype,
Ying Sheng's avatar
Ying Sheng committed
1371
1372
            choices=["auto", "half", "float16", "bfloat16", "float", "float32"],
            help="Data type for model weights and activations.\n\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
1373
            '* "auto" will use FP16 precision for FP32 and FP16 models, and '
Ying Sheng's avatar
Ying Sheng committed
1374
            "BF16 precision for BF16 models.\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
1375
1376
1377
1378
            '* "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
1379
1380
            '* "float32" for FP32 precision.',
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1381
1382
1383
1384
        parser.add_argument(
            "--quantization",
            type=str,
            default=ServerArgs.quantization,
1385
            choices=QUANTIZATION_CHOICES,
Lianmin Zheng's avatar
Lianmin Zheng committed
1386
1387
            help="The quantization method.",
        )
1388
1389
1390
1391
1392
1393
1394
1395
1396
        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. ",
        )
1397
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
1398
            "--kv-cache-dtype",
1399
            type=str,
Lianmin Zheng's avatar
Lianmin Zheng committed
1400
1401
1402
            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+.',
1403
        )
1404
1405
1406
1407
1408
        parser.add_argument(
            "--enable-fp32-lm-head",
            action="store_true",
            help="If set, the LM head outputs (logits) are in FP32.",
        )
1409

1410
        # Memory and scheduling
Lianmin Zheng's avatar
Lianmin Zheng committed
1411
1412
1413
1414
        parser.add_argument(
            "--mem-fraction-static",
            type=float,
            default=ServerArgs.mem_fraction_static,
1415
            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
1416
        )
1417
1418
1419
1420
1421
1422
        parser.add_argument(
            "--max-running-requests",
            type=int,
            default=ServerArgs.max_running_requests,
            help="The maximum number of running requests.",
        )
1423
1424
1425
1426
1427
1428
        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.",
        )
1429
1430
1431
1432
        parser.add_argument(
            "--max-total-tokens",
            type=int,
            default=ServerArgs.max_total_tokens,
1433
1434
            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.",
1435
        )
1436
1437
1438
1439
        parser.add_argument(
            "--chunked-prefill-size",
            type=int,
            default=ServerArgs.chunked_prefill_size,
1440
            help="The maximum number of tokens in a chunk for the chunked prefill. Setting this to -1 means disabling chunked prefill.",
1441
1442
1443
1444
1445
1446
1447
        )
        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
1448
        parser.add_argument(
1449
            "--schedule-policy",
Lianmin Zheng's avatar
Lianmin Zheng committed
1450
            type=str,
1451
            default=ServerArgs.schedule_policy,
1452
            choices=["lpm", "random", "fcfs", "dfs-weight", "lof", "priority"],
1453
            help="The scheduling policy of the requests.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1454
        )
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
        parser.add_argument(
            "--enable-priority-scheduling",
            action="store_true",
            default=ServerArgs.enable_priority_scheduling,
            help="Enable priority scheduling. Requests with higher priority integer values will be scheduled first by default.",
        )
        parser.add_argument(
            "--schedule-low-priority-values-first",
            action="store_true",
            default=ServerArgs.schedule_low_priority_values_first,
            help="If specified with --enable-priority-scheduling, the scheduler will schedule requests with lower priority integer values first.",
        )
        parser.add_argument(
            "--priority-scheduling-preemption-threshold",
            type=int,
            default=ServerArgs.priority_scheduling_preemption_threshold,
            help="Minimum difference in priorities for an incoming request to have to preempt running request(s).",
        )
1473
1474
1475
1476
        parser.add_argument(
            "--schedule-conservativeness",
            type=float,
            default=ServerArgs.schedule_conservativeness,
1477
            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.",
1478
        )
1479
1480
1481
1482
1483
1484
        parser.add_argument(
            "--page-size",
            type=int,
            default=ServerArgs.page_size,
            help="The number of tokens in a page.",
        )
tarinkk's avatar
tarinkk committed
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
        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
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
        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.",
        )
1509

Lianmin Zheng's avatar
Lianmin Zheng committed
1510
1511
1512
1513
1514
1515
1516
        # 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
1517
        parser.add_argument(
1518
            "--tensor-parallel-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
1519
            "--tp-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
1520
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
1521
            default=ServerArgs.tp_size,
1522
            help="The tensor parallelism size.",
1523
        )
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
        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.",
        )
1537
1538
1539
        parser.add_argument(
            "--stream-interval",
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
1540
            default=ServerArgs.stream_interval,
1541
            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",
1542
        )
1543
1544
1545
1546
1547
        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
1548
1549
1550
1551
        parser.add_argument(
            "--random-seed",
            type=int,
            default=ServerArgs.random_seed,
1552
            help="The random seed.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1553
        )
1554
1555
1556
1557
        parser.add_argument(
            "--constrained-json-whitespace-pattern",
            type=str,
            default=ServerArgs.constrained_json_whitespace_pattern,
Lianmin Zheng's avatar
Lianmin Zheng committed
1558
            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 ]*",
1559
        )
1560
1561
1562
1563
1564
1565
        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.",
        )
1566
1567
1568
1569
1570
1571
        parser.add_argument(
            "--dist-timeout",
            type=int,
            default=ServerArgs.dist_timeout,
            help="Set timeout for torch.distributed initialization.",
        )
1572
1573
1574
1575
        parser.add_argument(
            "--download-dir",
            type=str,
            default=ServerArgs.download_dir,
1576
            help="Model download directory for huggingface.",
1577
        )
1578
1579
1580
1581
1582
1583
        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.",
        )
1584
1585
1586
1587
1588
1589
        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,...",
        )
1590
1591
1592
1593
1594
        parser.add_argument(
            "--sleep-on-idle",
            action="store_true",
            help="Reduce CPU usage when sglang is idle.",
        )
1595
1596

        # Logging
Lianmin Zheng's avatar
Lianmin Zheng committed
1597
1598
1599
1600
        parser.add_argument(
            "--log-level",
            type=str,
            default=ServerArgs.log_level,
1601
            help="The logging level of all loggers.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1602
        )
1603
        parser.add_argument(
1604
1605
1606
1607
            "--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.",
1608
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1609
        parser.add_argument(
1610
            "--log-requests",
Lianmin Zheng's avatar
Lianmin Zheng committed
1611
            action="store_true",
1612
1613
1614
1615
1616
            help="Log metadata, inputs, outputs of all requests. The verbosity is decided by --log-requests-level",
        )
        parser.add_argument(
            "--log-requests-level",
            type=int,
1617
            default=ServerArgs.log_requests_level,
1618
1619
1620
1621
1622
1623
1624
1625
            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
1626
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1627
1628
1629
        parser.add_argument(
            "--show-time-cost",
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
1630
            help="Show time cost of custom marks.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1631
        )
1632
1633
1634
1635
1636
        parser.add_argument(
            "--enable-metrics",
            action="store_true",
            help="Enable log prometheus metrics.",
        )
1637
1638
1639
1640
1641
1642
1643
        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.",
        )
1644
1645
1646
1647
        parser.add_argument(
            "--tokenizer-metrics-custom-labels-header",
            type=str,
            default=ServerArgs.tokenizer_metrics_custom_labels_header,
1648
            help="Specify the HTTP header for passing custom labels for tokenizer metrics.",
1649
1650
        )
        parser.add_argument(
1651
            "--tokenizer-metrics-allowed-custom-labels",
1652
1653
            type=str,
            nargs="+",
1654
1655
            default=ServerArgs.tokenizer_metrics_allowed_custom_labels,
            help="The custom labels allowed for tokenizer metrics. The labels are specified via a dict in "
1656
            "'--tokenizer-metrics-custom-labels-header' field in HTTP requests, e.g., {'label1': 'value1', 'label2': "
1657
            "'value2'} is allowed if '--tokenizer-metrics-allowed-custom-labels label1 label2' is set.",
1658
        )
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
        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.",
        )
1686
1687
1688
        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 "
1689
1690
            "[984.0, 992.0, 996.0, 998.0, 1000.0, 1002.0, 1004.0, 1008.0, 1016.0]).); 'custom <value1> "
            "<value2> ...' uses custom bucket values (e.g., 'custom 10 50 100 500')."
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
        )
        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}",
        )
1706
1707
1708
1709
1710
1711
        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.",
        )
1712
1713
1714
1715
        parser.add_argument(
            "--decode-log-interval",
            type=int,
            default=ServerArgs.decode_log_interval,
1716
            help="The log interval of decode batch.",
1717
        )
1718
1719
1720
1721
1722
1723
        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
1724
1725
1726
1727
1728
        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.",
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
        )
        parser.add_argument(
            "--enable-trace",
            action="store_true",
            help="Enable opentelemetry trace",
        )
        parser.add_argument(
            "--oltp-traces-endpoint",
            type=str,
            default="localhost:4317",
            help="Config opentelemetry collector endpoint if --enable-trace is set. format: <ip>:<port>",
Lianmin Zheng's avatar
Lianmin Zheng committed
1740
        )
1741

1742
        # API related
Liangsheng Yin's avatar
Liangsheng Yin committed
1743
1744
1745
1746
        parser.add_argument(
            "--api-key",
            type=str,
            default=ServerArgs.api_key,
1747
            help="Set API key of the server. It is also used in the OpenAI API compatible server.",
Liangsheng Yin's avatar
Liangsheng Yin committed
1748
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1749
1750
1751
1752
1753
1754
        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.",
        )
1755
1756
1757
1758
1759
1760
        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
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
        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.",
        )
1773
        parser.add_argument(
1774
            "--file-storage-path",
1775
            type=str,
1776
            default=ServerArgs.file_storage_path,
1777
1778
            help="The path of the file storage in backend.",
        )
1779
1780
1781
1782
1783
        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
1784
1785
1786
1787
1788
1789
1790
        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())}.",
        )
1791
        tool_call_parser_choices = list(FunctionCallParser.ToolCallParserEnum.keys())
1792
1793
1794
        parser.add_argument(
            "--tool-call-parser",
            type=str,
1795
            choices=tool_call_parser_choices,
1796
            default=ServerArgs.tool_call_parser,
1797
            help=f"Specify the parser for handling tool-call interactions. Options include: {tool_call_parser_choices}.",
1798
        )
1799
1800
1801
1802
1803
1804
        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
1805

1806
1807
        # Data parallelism
        parser.add_argument(
1808
            "--data-parallel-size",
1809
1810
1811
            "--dp-size",
            type=int,
            default=ServerArgs.dp_size,
1812
            help="The data parallelism size.",
1813
1814
1815
1816
1817
        )
        parser.add_argument(
            "--load-balance-method",
            type=str,
            default=ServerArgs.load_balance_method,
1818
            help="The load balancing strategy for data parallelism.",
1819
1820
1821
            choices=[
                "round_robin",
                "shortest_queue",
1822
                "minimum_tokens",
1823
1824
            ],
        )
1825
1826
1827
1828
1829
1830
        parser.add_argument(
            "--load-watch-interval",
            type=float,
            default=ServerArgs.load_watch_interval,
            help="The interval of load watching in seconds.",
        )
1831
1832
1833
1834
1835
1836
        parser.add_argument(
            "--prefill-round-robin-balance",
            default=ServerArgs.prefill_round_robin_balance,
            action="store_true",
            help="Prefill is round robin balanced. This is used to promise decode server can get the correct dp rank.",
        )
1837

1838
        # Multi-node distributed serving
1839
        parser.add_argument(
1840
            "--dist-init-addr",
1841
            "--nccl-init-addr",  # For backward compatibility. This will be removed in the future.
1842
            type=str,
1843
            help="The host address for initializing distributed backend (e.g., `192.168.0.2:25000`).",
1844
1845
        )
        parser.add_argument(
Liangsheng Yin's avatar
Liangsheng Yin committed
1846
            "--nnodes", type=int, default=ServerArgs.nnodes, help="The number of nodes."
1847
        )
1848
1849
1850
        parser.add_argument(
            "--node-rank", type=int, default=ServerArgs.node_rank, help="The node rank."
        )
1851

Lianmin Zheng's avatar
Lianmin Zheng committed
1852
1853
1854
1855
1856
1857
1858
        # 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,
        )
1859
1860
1861
1862
1863
        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
1864

1865
        # LoRA
1866
1867
1868
1869
1870
1871
        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.",
        )
1872
1873
1874
1875
1876
1877
1878
1879
1880
        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,
1881
            choices=SUPPORTED_LORA_TARGET_MODULES + [LORA_TARGET_ALL_MODULES],
1882
1883
            nargs="*",
            default=None,
1884
1885
1886
            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.",
1887
        )
1888
1889
1890
1891
1892
1893
        parser.add_argument(
            "--lora-paths",
            type=str,
            nargs="*",
            default=None,
            action=LoRAPathAction,
1894
            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}',
1895
1896
1897
1898
1899
        )
        parser.add_argument(
            "--max-loras-per-batch",
            type=int,
            default=8,
1900
1901
            help="Maximum number of adapters for a running batch, include base-only request.",
        )
1902
1903
1904
1905
1906
1907
        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`.",
        )
1908
1909
1910
        parser.add_argument(
            "--lora-backend",
            type=str,
1911
1912
            choices=LORA_BACKEND_CHOICES,
            default=ServerArgs.lora_backend,
1913
            help="Choose the kernel backend for multi-LoRA serving.",
1914
        )
1915
1916
1917
1918
1919
1920
1921
        parser.add_argument(
            "--max-lora-chunk-size",
            type=int,
            default=ServerArgs.max_lora_chunk_size,
            choices=[16, 32, 64, 128],
            help="Maximum chunk size for the ChunkedSGMV LoRA backend. Only used when --lora-backend is 'csgmv'. Choosing a larger value might improve performance.",
        )
1922
1923

        # Kernel backend
1924
1925
1926
        parser.add_argument(
            "--attention-backend",
            type=str,
1927
            choices=ATTENTION_BACKEND_CHOICES,
1928
1929
1930
            default=ServerArgs.attention_backend,
            help="Choose the kernels for attention layers.",
        )
1931
1932
1933
        parser.add_argument(
            "--prefill-attention-backend",
            type=str,
1934
            choices=ATTENTION_BACKEND_CHOICES,
1935
1936
1937
            default=ServerArgs.prefill_attention_backend,
            help="Choose the kernels for prefill attention layers (have priority over --attention-backend).",
        )
1938
1939
1940
        parser.add_argument(
            "--decode-attention-backend",
            type=str,
1941
            choices=ATTENTION_BACKEND_CHOICES,
1942
1943
1944
            default=ServerArgs.decode_attention_backend,
            help="Choose the kernels for decode attention layers (have priority over --attention-backend).",
        )
1945
1946
1947
1948
1949
1950
1951
        parser.add_argument(
            "--sampling-backend",
            type=str,
            choices=["flashinfer", "pytorch"],
            default=ServerArgs.sampling_backend,
            help="Choose the kernels for sampling layers.",
        )
1952
1953
1954
        parser.add_argument(
            "--grammar-backend",
            type=str,
1955
            choices=GRAMMAR_BACKEND_CHOICES,
1956
            default=ServerArgs.grammar_backend,
Lianmin Zheng's avatar
Lianmin Zheng committed
1957
            help="Choose the backend for grammar-guided decoding.",
1958
        )
1959
1960
1961
        parser.add_argument(
            "--mm-attention-backend",
            type=str,
1962
            choices=["sdpa", "fa3", "triton_attn", "ascend_attn"],
1963
1964
1965
            default=ServerArgs.mm_attention_backend,
            help="Set multimodal attention backend.",
        )
1966

1967
1968
1969
1970
        # Speculative decoding
        parser.add_argument(
            "--speculative-algorithm",
            type=str,
1971
            choices=["EAGLE", "EAGLE3", "NEXTN", "STANDALONE", "NGRAM"],
1972
1973
1974
1975
            help="Speculative algorithm.",
        )
        parser.add_argument(
            "--speculative-draft-model-path",
1976
            "--speculative-draft-model",
1977
1978
1979
            type=str,
            help="The path of the draft model weights. This can be a local folder or a Hugging Face repo ID.",
        )
1980
1981
1982
1983
1984
1985
1986
1987
        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.",
        )
1988
1989
1990
1991
1992
1993
1994
1995
1996
        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,
1997
            help="The number of tokens sampled from the draft model in eagle2 each step.",
1998
1999
            default=ServerArgs.speculative_eagle_topk,
        )
2000
2001
2002
        parser.add_argument(
            "--speculative-num-draft-tokens",
            type=int,
2003
            help="The number of tokens sampled from the draft model in Speculative Decoding.",
2004
2005
            default=ServerArgs.speculative_num_draft_tokens,
        )
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
        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,
        )
2018
2019
2020
2021
2022
2023
        parser.add_argument(
            "--speculative-token-map",
            type=str,
            help="The path of the draft model's small vocab table.",
            default=ServerArgs.speculative_token_map,
        )
2024
        parser.add_argument(
2025
            "--speculative-attention-mode",
2026
2027
            type=str,
            choices=["prefill", "decode"],
2028
2029
            help="Attention backend for speculative decoding operations (both target verify and draft extend). Can be one of 'prefill' (default) or 'decode'.",
            default=ServerArgs.speculative_attention_mode,
2030
        )
2031
        # Ngram speculative decoding
2032
        parser.add_argument(
2033
            "--speculative-ngram-min-match-window-size",
2034
            type=int,
2035
2036
            default=ServerArgs.speculative_ngram_min_match_window_size,
            help="The minimum window size for pattern matching in ngram speculative decoding.",
2037
2038
        )
        parser.add_argument(
2039
            "--speculative-ngram-max-match-window-size",
2040
            type=int,
2041
2042
            default=ServerArgs.speculative_ngram_max_match_window_size,
            help="The maximum window size for pattern matching in ngram speculative decoding.",
2043
2044
        )
        parser.add_argument(
2045
            "--speculative-ngram-min-bfs-breadth",
2046
            type=int,
2047
2048
            default=ServerArgs.speculative_ngram_min_bfs_breadth,
            help="The minimum breadth for BFS (Breadth-First Search) in ngram speculative decoding.",
2049
2050
        )
        parser.add_argument(
2051
            "--speculative-ngram-max-bfs-breadth",
2052
            type=int,
2053
2054
            default=ServerArgs.speculative_ngram_max_bfs_breadth,
            help="The maximum breadth for BFS (Breadth-First Search) in ngram speculative decoding.",
2055
2056
        )
        parser.add_argument(
2057
            "--speculative-ngram-match-type",
2058
2059
            type=str,
            choices=["BFS", "PROB"],
2060
            default=ServerArgs.speculative_ngram_match_type,
2061
2062
2063
            help="The match type for cache tree.",
        )
        parser.add_argument(
2064
            "--speculative-ngram-branch-length",
2065
            type=int,
2066
2067
            default=ServerArgs.speculative_ngram_branch_length,
            help="The branch length for ngram speculative decoding.",
2068
2069
        )
        parser.add_argument(
2070
            "--speculative-ngram-capacity",
2071
            type=int,
2072
2073
            default=ServerArgs.speculative_ngram_capacity,
            help="The cache capacity for ngram speculative decoding.",
2074
        )
2075
2076
2077
2078
2079

        # Expert parallelism
        parser.add_argument(
            "--expert-parallel-size",
            "--ep-size",
Cheng Wan's avatar
Cheng Wan committed
2080
            "--ep",
2081
2082
2083
2084
2085
            type=int,
            default=ServerArgs.ep_size,
            help="The expert parallelism size.",
        )
        parser.add_argument(
2086
2087
            "--moe-a2a-backend",
            type=str,
2088
            choices=["none", "deepep"],
2089
2090
            default=ServerArgs.moe_a2a_backend,
            help="Choose the backend for MoE A2A.",
2091
        )
2092
        parser.add_argument(
2093
2094
2095
2096
2097
2098
2099
2100
            "--moe-runner-backend",
            type=str,
            choices=[
                "auto",
                "triton",
                "triton_kernel",
                "flashinfer_trtllm",
                "flashinfer_cutlass",
2101
                "flashinfer_mxfp4",
2102
                "flashinfer_cutedsl",
2103
2104
2105
            ],
            default=ServerArgs.moe_runner_backend,
            help="Choose the runner backend for MoE.",
2106
2107
        )
        parser.add_argument(
2108
2109
            "--flashinfer-mxfp4-moe-precision",
            type=str,
2110
            choices=["default", "bf16"],
2111
2112
2113
2114
            default=ServerArgs.flashinfer_mxfp4_moe_precision,
            help="Choose the computation precision of flashinfer mxfp4 moe",
        )
        parser.add_argument(
2115
2116
            "--enable-flashinfer-allreduce-fusion",
            action="store_true",
2117
            help="Enable FlashInfer allreduce fusion with Residual RMSNorm.",
2118
        )
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
        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.",
        )
2167
2168
2169
2170
2171
2172
        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].",
        )
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
        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.",
        )
2202

Yi Zhang's avatar
Yi Zhang committed
2203
2204
2205
2206
2207
        # Mamba Cache
        parser.add_argument(
            "--max-mamba-cache-size",
            type=int,
            default=ServerArgs.max_mamba_cache_size,
2208
            help="The maximum size of the mamba cache.",
Yi Zhang's avatar
Yi Zhang committed
2209
2210
2211
2212
2213
2214
        )
        parser.add_argument(
            "--mamba-ssm-dtype",
            type=str,
            default=ServerArgs.mamba_ssm_dtype,
            choices=["float32", "bfloat16"],
2215
            help="The data type of the SSM states in mamba cache.",
Yi Zhang's avatar
Yi Zhang committed
2216
        )
2217

Lianmin Zheng's avatar
Lianmin Zheng committed
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
        # 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.",
        )
2243
2244
2245
2246
2247
2248
2249
        parser.add_argument(
            "--radix-eviction-policy",
            type=str,
            choices=["lru", "lfu"],
            default=ServerArgs.radix_eviction_policy,
            help="The eviction policy of radix trees. 'lru' stands for Least Recently Used, 'lfu' stands for Least Frequently Used.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2250
2251
2252
2253
2254
2255
2256
        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",
        )
2257
2258
2259
        parser.add_argument(
            "--hicache-mem-layout",
            type=str,
2260
            choices=["layer_first", "page_first", "page_first_direct"],
2261
2262
2263
            default=ServerArgs.hicache_mem_layout,
            help="The layout of host memory pool for hierarchical cache.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2264
2265
2266
        parser.add_argument(
            "--hicache-storage-backend",
            type=str,
2267
            choices=["file", "mooncake", "hf3fs", "nixl", "aibrix", "dynamic", "eic"],
Lianmin Zheng's avatar
Lianmin Zheng committed
2268
            default=ServerArgs.hicache_storage_backend,
2269
2270
2271
2272
            help="The storage backend for hierarchical KV cache. "
            "Built-in backends: file, mooncake, hf3fs, nixl, aibrix. "
            "For dynamic backend, use --hicache-storage-backend-extra-config to specify: "
            "backend_name (custom name), module_path (Python module path), class_name (backend class name).",
Lianmin Zheng's avatar
Lianmin Zheng committed
2273
        )
pansicheng's avatar
pansicheng committed
2274
2275
2276
2277
2278
2279
2280
        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.",
        )
2281
2282
2283
2284
2285
2286
        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.",
        )
2287
2288
2289
2290
2291
2292
        # 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
2293

2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
        # 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
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
        # 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.",
        )

2363
        # Optimization/debug options
Liangsheng Yin's avatar
Liangsheng Yin committed
2364
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
2365
            "--disable-radix-cache",
Liangsheng Yin's avatar
Liangsheng Yin committed
2366
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
2367
            help="Disable RadixAttention for prefix caching.",
Liangsheng Yin's avatar
Liangsheng Yin committed
2368
        )
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
        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.",
        )
2381
2382
2383
        parser.add_argument(
            "--disable-cuda-graph",
            action="store_true",
2384
            help="Disable cuda graph.",
2385
        )
2386
        parser.add_argument(
2387
2388
            "--disable-cuda-graph-padding",
            action="store_true",
2389
            help="Disable cuda graph when padding is needed. Still uses cuda graph when padding is not needed.",
2390
        )
2391
2392
2393
2394
2395
        parser.add_argument(
            "--enable-profile-cuda-graph",
            action="store_true",
            help="Enable profiling of cuda graph capture.",
        )
2396
2397
2398
2399
2400
        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.",
        )
2401
2402
2403
2404
2405
        parser.add_argument(
            "--enable-nccl-nvls",
            action="store_true",
            help="Enable NCCL NVLS for prefill heavy requests when available.",
        )
2406
2407
2408
2409
2410
        parser.add_argument(
            "--enable-symm-mem",
            action="store_true",
            help="Enable NCCL symmetric memory for fast collectives.",
        )
2411
2412
2413
2414
2415
        parser.add_argument(
            "--disable-flashinfer-cutlass-moe-fp4-allgather",
            action="store_true",
            help="Disables quantize before all-gather for flashinfer cutlass moe.",
        )
2416
2417
2418
2419
2420
        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.",
        )
2421
        parser.add_argument(
2422
            "--disable-outlines-disk-cache",
2423
            action="store_true",
2424
            help="Disable disk cache of outlines to avoid possible crashes related to file system or high concurrency.",
2425
        )
2426
2427
2428
2429
2430
        parser.add_argument(
            "--disable-custom-all-reduce",
            action="store_true",
            help="Disable the custom all-reduce kernel and fall back to NCCL.",
        )
2431
2432
2433
2434
2435
        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
2436
        parser.add_argument(
2437
            "--disable-overlap-schedule",
Lianmin Zheng's avatar
Lianmin Zheng committed
2438
            action="store_true",
2439
            help="Disable the overlap scheduler, which overlaps the CPU scheduler with GPU model worker.",
Lianmin Zheng's avatar
Lianmin Zheng committed
2440
        )
2441
2442
2443
        parser.add_argument(
            "--enable-mixed-chunk",
            action="store_true",
2444
            help="Enabling mixing prefill and decode in a batch when using chunked prefill.",
2445
        )
Ke Bao's avatar
Ke Bao committed
2446
2447
2448
        parser.add_argument(
            "--enable-dp-attention",
            action="store_true",
2449
            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
2450
        )
2451
2452
2453
2454
2455
        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.",
        )
2456
2457
2458
2459
2460
        parser.add_argument(
            "--enable-two-batch-overlap",
            action="store_true",
            help="Enabling two micro batches to overlap.",
        )
2461
2462
2463
2464
2465
        parser.add_argument(
            "--enable-single-batch-overlap",
            action="store_true",
            help="Let computation and communication overlap within one micro batch.",
        )
2466
2467
2468
2469
2470
2471
        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.",
        )
2472
2473
2474
        parser.add_argument(
            "--enable-torch-compile",
            action="store_true",
2475
2476
            help="Optimize the model with torch.compile. Experimental feature.",
        )
2477
        parser.add_argument(
2478
            "--torch-compile-max-bs",
2479
            type=int,
2480
            default=ServerArgs.torch_compile_max_bs,
2481
2482
            help="Set the maximum batch size when using torch compile.",
        )
2483
2484
2485
2486
        parser.add_argument(
            "--torchao-config",
            type=str,
            default=ServerArgs.torchao_config,
2487
            help="Optimize the model with torchao. Experimental feature. Current choices are: int8dq, int8wo, int4wo-<group_size>, fp8wo, fp8dq-per_tensor, fp8dq-per_row",
2488
        )
2489
2490
2491
2492
2493
        parser.add_argument(
            "--enable-nan-detection",
            action="store_true",
            help="Enable the NaN detection for debugging purposes.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2494
        parser.add_argument(
2495
            "--enable-p2p-check",
Lianmin Zheng's avatar
Lianmin Zheng committed
2496
            action="store_true",
2497
            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
Lianmin Zheng's avatar
Lianmin Zheng committed
2498
        )
2499
        parser.add_argument(
2500
            "--triton-attention-reduce-in-fp32",
2501
            action="store_true",
2502
            help="Cast the intermediate attention results to fp32 to avoid possible crashes related to fp16."
2503
            "This only affects Triton attention kernels.",
2504
        )
2505
2506
2507
2508
2509
2510
        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.",
        )
2511
2512
2513
2514
2515
2516
        parser.add_argument(
            "--triton-attention-split-tile-size",
            type=int,
            default=ServerArgs.triton_attention_split_tile_size,
            help="The size of split KV tile in flash decoding Triton kernel. Used for deterministic inference.",
        )
2517
2518
2519
2520
2521
2522
2523
2524
        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.",
        )
2525
2526
2527
2528
2529
        parser.add_argument(
            "--delete-ckpt-after-loading",
            action="store_true",
            help="Delete the model checkpoint after loading the model.",
        )
2530
2531
2532
2533
2534
        parser.add_argument(
            "--enable-memory-saver",
            action="store_true",
            help="Allow saving memory using release_memory_occupation and resume_memory_occupation",
        )
2535
2536
2537
2538
2539
        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.",
        )
2540
2541
2542
2543
2544
        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)",
        )
2545
        parser.add_argument(
2546
            "--flashinfer-mla-disable-ragged",
2547
            action="store_true",
2548
            help="Not using ragged prefill wrapper when running flashinfer mla",
2549
        )
2550
        parser.add_argument(
2551
2552
2553
            "--disable-shared-experts-fusion",
            action="store_true",
            help="Disable shared experts fusion optimization for deepseek v3/r1.",
2554
        )
2555
2556
2557
2558
2559
        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
2560
2561
2562
2563
2564
        parser.add_argument(
            "--disable-fast-image-processor",
            action="store_true",
            help="Adopt base image processor instead of fast image processor.",
        )
2565
2566
2567
2568
2569
        parser.add_argument(
            "--keep-mm-feature-on-device",
            action="store_true",
            help="Keep multimodal feature tensors on device after processing to save D2H copy.",
        )
2570
2571
2572
2573
2574
        parser.add_argument(
            "--enable-return-hidden-states",
            action="store_true",
            help="Enable returning hidden states with responses.",
        )
2575
2576
2577
2578
2579
2580
        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.",
        )
2581
2582
2583
2584
2585
2586
        parser.add_argument(
            "--numa-node",
            type=int,
            nargs="+",
            help="Sets the numa node for the subprocesses. i-th element corresponds to i-th subprocess.",
        )
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606

        # 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.",
        )
2607
2608
2609
2610
2611
        parser.add_argument(
            "--debug-tensor-dump-prefill-only",
            action="store_true",
            help="Only dump the tensors for prefill requests (i.e. batch size > 1).",
        )
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
        parser.add_argument(
            "--enable-dynamic-batch-tokenizer",
            action="store_true",
            help="Enable async dynamic batch tokenizer for improved performance when multiple requests arrive concurrently.",
        )
        parser.add_argument(
            "--dynamic-batch-tokenizer-batch-size",
            type=int,
            default=ServerArgs.dynamic_batch_tokenizer_batch_size,
            help="[Only used if --enable-dynamic-batch-tokenizer is set] Maximum batch size for dynamic batch tokenizer.",
        )
        parser.add_argument(
            "--dynamic-batch-tokenizer-batch-timeout",
            type=float,
            default=ServerArgs.dynamic_batch_tokenizer_batch_timeout,
            help="[Only used if --enable-dynamic-batch-tokenizer is set] Timeout in seconds for batching tokenization requests.",
        )
2629

Lianmin Zheng's avatar
Lianmin Zheng committed
2630
        # PD disaggregation
Byron Hsu's avatar
Byron Hsu committed
2631
2632
2633
        parser.add_argument(
            "--disaggregation-mode",
            type=str,
2634
            default=ServerArgs.disaggregation_mode,
Byron Hsu's avatar
Byron Hsu committed
2635
2636
2637
            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',
        )
2638
2639
2640
2641
        parser.add_argument(
            "--disaggregation-transfer-backend",
            type=str,
            default=ServerArgs.disaggregation_transfer_backend,
2642
            choices=DISAGG_TRANSFER_BACKEND_CHOICES,
2643
2644
            help="The backend for disaggregation transfer. Default is mooncake.",
        )
2645
2646
2647
2648
2649
2650
        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
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
        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.",
        )
2669
2670
2671
2672
        parser.add_argument(
            "--disaggregation-ib-device",
            type=str,
            default=ServerArgs.disaggregation_ib_device,
2673
2674
2675
            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.",
2676
        )
2677
2678
2679
2680
2681
        parser.add_argument(
            "--disaggregation-decode-enable-offload-kvcache",
            action="store_true",
            help="Enable async KV cache offloading on decode server (PD mode).",
        )
2682
2683
2684
2685
2686
2687
        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.",
        )
2688
2689
2690
2691
2692
2693
        parser.add_argument(
            "--disaggregation-decode-polling-interval",
            type=int,
            default=ServerArgs.disaggregation_decode_polling_interval,
            help="The interval to poll requests in decode server. Can be set to >1 to reduce the overhead of this.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2694
2695

        # Custom weight loader
2696
2697
2698
2699
2700
2701
2702
        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",
        )
2703
2704
2705
2706
2707
        parser.add_argument(
            "--weight-loader-disable-mmap",
            action="store_true",
            help="Disable mmap while loading weight using safetensors.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
        parser.add_argument(
            "--remote-instance-weight-loader-seed-instance-ip",
            type=str,
            default=ServerArgs.remote_instance_weight_loader_seed_instance_ip,
            help="The ip of the seed instance for loading weights from remote instance.",
        )
        parser.add_argument(
            "--remote-instance-weight-loader-seed-instance-service-port",
            type=int,
            default=ServerArgs.remote_instance_weight_loader_seed_instance_service_port,
            help="The service port of the seed instance for loading weights from remote instance.",
        )
        parser.add_argument(
            "--remote-instance-weight-loader-send-weights-group-ports",
            type=json_list_type,
            default=ServerArgs.remote_instance_weight_loader_send_weights_group_ports,
            help="The communication group ports for loading weights from remote instance.",
        )
2726
2727

        # For PD-Multiplexing
2728
2729
2730
2731
2732
        parser.add_argument(
            "--enable-pdmux",
            action="store_true",
            help="Enable PD-Multiplexing, PD running on greenctx stream.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2733

2734
2735
2736
2737
2738
2739
        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
2740

2741
2742
2743
2744
2745
2746
2747
        # For deterministic inference
        parser.add_argument(
            "--enable-deterministic-inference",
            action="store_true",
            help="Enable deterministic inference mode with batch invariant ops.",
        )

2748
2749
2750
        # Deprecated arguments
        parser.add_argument(
            "--enable-ep-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2751
2752
            action=DeprecatedAction,
            help="NOTE: --enable-ep-moe is deprecated. Please set `--ep-size` to the same value as `--tp-size` instead.",
2753
2754
2755
        )
        parser.add_argument(
            "--enable-deepep-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2756
2757
            action=DeprecatedAction,
            help="NOTE: --enable-deepep-moe is deprecated. Please set `--moe-a2a-backend` to 'deepep' instead.",
2758
        )
2759
2760
        parser.add_argument(
            "--enable-flashinfer-cutlass-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2761
2762
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-cutlass-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_cutlass' instead.",
2763
        )
2764
2765
        parser.add_argument(
            "--enable-flashinfer-cutedsl-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2766
2767
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-cutedsl-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_cutedsl' instead.",
2768
        )
2769
2770
        parser.add_argument(
            "--enable-flashinfer-trtllm-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2771
2772
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-trtllm-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_trtllm' instead.",
2773
2774
2775
        )
        parser.add_argument(
            "--enable-triton-kernel-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2776
2777
            action=DeprecatedAction,
            help="NOTE: --enable-triton-kernel-moe is deprecated. Please set `--moe-runner-backend` to 'triton_kernel' instead.",
2778
        )
2779
2780
        parser.add_argument(
            "--enable-flashinfer-mxfp4-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2781
2782
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-mxfp4-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_mxfp4' instead.",
2783
        )
2784

2785
2786
2787
2788
2789
2790
2791
        # Configuration file support
        parser.add_argument(
            "--config",
            type=str,
            help="Read CLI options from a config file. Must be a YAML file with configuration options.",
        )

Lianmin Zheng's avatar
Lianmin Zheng committed
2792
2793
    @classmethod
    def from_cli_args(cls, args: argparse.Namespace):
2794
        args.tp_size = args.tensor_parallel_size
2795
        args.pp_size = args.pipeline_parallel_size
2796
        args.dp_size = args.data_parallel_size
xiaobochen's avatar
xiaobochen committed
2797
        args.ep_size = args.expert_parallel_size
2798

Lianmin Zheng's avatar
Lianmin Zheng committed
2799
2800
2801
2802
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        return cls(**{attr: getattr(args, attr) for attr in attrs})

    def url(self):
2803
        if is_valid_ipv6_address(self.host):
2804
2805
2806
            return f"http://[{self.host}]:{self.port}"
        else:
            return f"http://{self.host}:{self.port}"
Lianmin Zheng's avatar
Lianmin Zheng committed
2807

Lianmin Zheng's avatar
Lianmin Zheng committed
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
    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

2819
    def check_server_args(self):
2820
        # Check parallel size constraints
2821
        assert (
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
            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."

2832
        assert not (
2833
2834
            self.dp_size > 1 and self.nnodes != 1 and not self.enable_dp_attention
        ), "multi-node data parallel is not supported unless dp attention!"
2835

2836
        assert self.base_gpu_id >= 0, "base_gpu_id must be non-negative"
2837
        assert self.gpu_id_step >= 1, "gpu_id_step must be positive"
2838

Lianmin Zheng's avatar
Lianmin Zheng committed
2839
2840
2841
2842
2843
        assert self.moe_dense_tp_size in {
            1,
            None,
        }, "moe_dense_tp_size only support 1 and None currently"

2844
        # Check LoRA
2845
2846
        self.check_lora_server_args()

2847
2848
2849
2850
2851
2852
2853
        # 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
2854
        # Skip validation if chunked prefill is disabled (i.e., size <= 0).
2855
2856
        # Skip validation if disaggregation mode is decode.
        if self.chunked_prefill_size > 0 and self.disaggregation_mode != "decode":
2857
2858
2859
            assert (
                self.chunked_prefill_size % self.page_size == 0
            ), "chunked_prefill_size must be divisible by page_size"
2860

2861
2862
        # Check multi tokenizer
        assert self.tokenizer_worker_num > 0, "Tokenizer worker num must >= 1"
2863
2864
2865
2866
2867
2868
        self.validate_buckets_rule(
            "--prompt-tokens-buckets", self.prompt_tokens_buckets
        )
        self.validate_buckets_rule(
            "--generation-tokens-buckets", self.generation_tokens_buckets
        )
2869

2870
2871
2872
2873
2874
2875
2876
        # Check scheduling policy
        if self.enable_priority_scheduling:
            assert self.schedule_policy in [
                "fcfs",
                "lof",
            ], f"To use priority scheduling, schedule_policy must be 'fcfs' or 'lof'. '{self.schedule_policy}' is not supported."

2877
    def check_lora_server_args(self):
2878
        assert self.max_loras_per_batch > 0, "max_loras_per_batch must be positive"
2879

2880
2881
2882
2883
        # 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
2884
                logger.warning(
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
                    "--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
2895
                self.lora_paths = []
2896
                for lora_path in lora_paths:
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
                    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),
2915
                        )
2916
                    else:
2917
2918
2919
                        raise ValueError(
                            f"Invalid type for item in --lora-paths list: {type(lora_path)}. "
                            "Expected a string or a dictionary."
2920
                        )
2921
                    self.lora_paths.append(lora_ref)
2922
            elif isinstance(self.lora_paths, dict):
2923
2924
                self.lora_paths = [
                    LoRARef(lora_name=k, lora_path=v, pinned=False)
2925
                    for k, v in self.lora_paths.items()
2926
                ]
2927
            elif self.lora_paths is None:
2928
                self.lora_paths = []
2929
2930
2931
2932
2933
            else:
                raise ValueError(
                    f"Invalid type for --lora-paths: {type(self.lora_paths)}. "
                    "Expected a list or a dictionary."
                )
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947

            # 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."
2948

2949
2950
2951
2952
2953
2954
            # 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}"
                )
2955
                assert len(self.lora_paths) <= self.max_loaded_loras, (
2956
2957
2958
2959
                    "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)}"
                )

2960
2961
2962
2963
2964
2965
            if self.max_lora_chunk_size is not None:
                assert (
                    16 <= self.max_lora_chunk_size <= 128
                    and (self.max_lora_chunk_size & (self.max_lora_chunk_size - 1)) == 0
                ), "--max-lora-chunk-size must be a power of 2 between 16 and 128."

Lianmin Zheng's avatar
Lianmin Zheng committed
2966
2967
2968
2969
2970
2971
2972
2973
    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}"
        )

2974
2975
2976
2977
2978
2979
2980
2981
2982
    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",
2983
2984
            "custom",
        ], f"Unsupported {arg_name} rule type: '{rule}'. Must be one of: 'tse', 'default', 'custom'"
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006

        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)}"

3007
        elif rule == "custom":
3008
3009
            assert (
                len(buckets_rule) >= 2
3010
            ), f"{arg_name} custom rule requires at least one bucket value: ['custom', value1, ...]"
3011
3012
3013
            try:
                bucket_values = [float(x) for x in buckets_rule[1:]]
            except ValueError:
3014
                assert False, f"{arg_name} custom rule bucket values must be numeric"
3015
3016
            assert len(set(bucket_values)) == len(
                bucket_values
3017
            ), f"{arg_name} custom rule bucket values should not contain duplicates"
3018
3019
            assert all(
                val >= 0 for val in bucket_values
3020
            ), f"{arg_name} custom rule bucket values should be non-negative"
3021

Lianmin Zheng's avatar
Lianmin Zheng committed
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
    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
3060

Lianmin Zheng's avatar
Lianmin Zheng committed
3061
def prepare_server_args(argv: List[str]) -> ServerArgs:
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
    """
    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.
    """
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
    # Import here to avoid circular imports
    from sglang.srt.server_args_config_parser import ConfigArgumentMerger

    # Check for config file and merge arguments if present
    if "--config" in argv:
        # Extract boolean actions from the parser to handle them correctly
        parser = argparse.ArgumentParser()
        ServerArgs.add_cli_args(parser)

        # Get boolean action destinations
        boolean_actions = []
        for action in parser._actions:
            if hasattr(action, "dest") and hasattr(action, "action"):
                if action.action in ["store_true", "store_false"]:
                    boolean_actions.append(action.dest)

        # Merge config file arguments with CLI arguments
        config_merger = ConfigArgumentMerger(boolean_actions=boolean_actions)
        argv = config_merger.merge_config_with_args(argv)

3092
3093
    parser = argparse.ArgumentParser()
    ServerArgs.add_cli_args(parser)
Lianmin Zheng's avatar
Lianmin Zheng committed
3094
    raw_args = parser.parse_args(argv)
3095
3096
3097
3098
    server_args = ServerArgs.from_cli_args(raw_args)
    return server_args


3099
3100
3101
ZMQ_TCP_PORT_DELTA = 233


Lianmin Zheng's avatar
Lianmin Zheng committed
3102
3103
@dataclasses.dataclass
class PortArgs:
3104
3105
3106
3107
3108
3109
    # 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
3110

3111
3112
    # The port for nccl initialization (torch.dist)
    nccl_port: int
3113

3114
3115
3116
    # The ipc filename for rpc call between Engine and Scheduler
    rpc_ipc_name: str

3117
3118
3119
    # The ipc filename for Scheduler to send metrics
    metrics_ipc_name: str

3120
3121
3122
    # The ipc filename for Tokenizer and worker tokenizer
    tokenizer_worker_ipc_name: Optional[str]

3123
    @staticmethod
3124
    def init_new(server_args, dp_rank: Optional[int] = None) -> "PortArgs":
3125
        if server_args.nccl_port is None:
Lianmin Zheng's avatar
Lianmin Zheng committed
3126
            nccl_port = server_args.port + random.randint(100, 1000)
3127
            while True:
Lianmin Zheng's avatar
Lianmin Zheng committed
3128
                if is_port_available(nccl_port):
3129
                    break
Lianmin Zheng's avatar
Lianmin Zheng committed
3130
3131
                if nccl_port < 60000:
                    nccl_port += 42
3132
                else:
Lianmin Zheng's avatar
Lianmin Zheng committed
3133
                    nccl_port -= 43
3134
        else:
Lianmin Zheng's avatar
Lianmin Zheng committed
3135
            nccl_port = server_args.nccl_port
3136

3137
3138
3139
3140
3141
3142
        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
3143
                nccl_port=nccl_port,
3144
                rpc_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
3145
                metrics_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
3146
                tokenizer_worker_ipc_name=None,
3147
3148
3149
3150
3151
            )
        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
3152
3153
3154
            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))
3155
3156
            else:
                dist_init_addr = server_args.dist_init_addr.split(":")
Vincent's avatar
Vincent committed
3157

3158
3159
3160
3161
3162
3163
            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
3164
3165
3166
            detokenizer_port = port_base + 1
            rpc_port = port_base + 2
            metrics_ipc_name = port_base + 3
3167
            if dp_rank is None:
3168
                # TokenizerManager to DataParallelController
3169
                scheduler_input_port = port_base + 4
3170
            else:
3171
                scheduler_input_port = port_base + 4 + 1 + dp_rank
3172
3173
3174
3175

            return PortArgs(
                tokenizer_ipc_name=f"tcp://{dist_init_host}:{port_base}",
                scheduler_input_ipc_name=f"tcp://{dist_init_host}:{scheduler_input_port}",
3176
                detokenizer_ipc_name=f"tcp://{dist_init_host}:{detokenizer_port}",
Lianmin Zheng's avatar
Lianmin Zheng committed
3177
                nccl_port=nccl_port,
3178
3179
                rpc_ipc_name=f"tcp://{dist_init_host}:{rpc_port}",
                metrics_ipc_name=f"tcp://{dist_init_host}:{metrics_ipc_name}",
3180
                tokenizer_worker_ipc_name=None,
3181
            )
3182

3183
3184
3185

class LoRAPathAction(argparse.Action):
    def __call__(self, parser, namespace, values, option_string=None):
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
        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)
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211


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)
3212
3213


3214
3215
3216
3217
def print_deprecated_warning(message: str):
    logger.warning(f"\033[33m{message}\033[0m")


3218
def auto_choose_speculative_params(self: ServerArgs):
3219
3220
3221
3222
3223
    """
    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
3224
    hf_config = self.get_hf_config()
3225
    arch = hf_config.architectures[0]
3226
3227
3228
    if self.speculative_algorithm == "STANDALONE":
        # The default value for standalone speculative decoding
        return (3, 1, 4)
3229
3230
3231
    if arch in ["LlamaForCausalLM"]:
        # The default value for llama
        return (5, 4, 8)
3232
3233
3234
3235
    elif arch in [
        "DeepseekV3ForCausalLM",
        "DeepseekV2ForCausalLM",
        "GptOssForCausalLM",
strgrb's avatar
strgrb committed
3236
3237
        "BailingMoeForCausalLM",
        "BailingMoeV2ForCausalLM",
3238
3239
    ]:
        # The default value for deepseek and gpt-oss
3240
        return (3, 1, 4)
3241
3242
3243
3244
3245
    elif arch in ["Grok1ForCausalLM", "Grok1VForCausalLM"]:
        return (5, 4, 8)
    else:
        # The default value for all other models
        return (5, 4, 8)