server_args.py 129 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
    "flashinfer",
    "flashmla",
    "trtllm_mla",
    "trtllm_mha",
    "dual_chunk_flash_attn",
Yi Zhang's avatar
Yi Zhang committed
103
    "hybrid_linear_attn",
104
105
106
107
108
109
110
111
    # AMD specific
    "aiter",
    "wave",
    # Other platforms
    "intel_amx",
    "ascend",
]

112
113
LORA_BACKEND_CHOICES = ["triton", "csgmv"]

114
115
DISAGG_TRANSFER_BACKEND_CHOICES = ["mooncake", "nixl", "ascend", "fake"]

116
117
GRAMMAR_BACKEND_CHOICES = ["xgrammar", "outlines", "llguidance", "none"]

118
DETERMINISTIC_ATTENTION_BACKEND_CHOICES = ["flashinfer", "fa3", "triton"]
119

120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137

# 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)


138
139
140
141
def add_grammar_backend_choices(choices):
    GRAMMAR_BACKEND_CHOICES.extend(choices)


Lianmin Zheng's avatar
Lianmin Zheng committed
142
143
@dataclasses.dataclass
class ServerArgs:
Lianmin Zheng's avatar
Lianmin Zheng committed
144
    # Model and tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
145
146
147
    model_path: str
    tokenizer_path: Optional[str] = None
    tokenizer_mode: str = "auto"
148
    tokenizer_worker_num: int = 1
149
    skip_tokenizer_init: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
150
    load_format: str = "auto"
151
    model_loader_extra_config: str = "{}"
152
    trust_remote_code: bool = False
153
    context_length: Optional[int] = None
154
    is_embedding: bool = False
155
    enable_multimodal: Optional[bool] = None
156
    revision: Optional[str] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
157
    model_impl: str = "auto"
Lianmin Zheng's avatar
Lianmin Zheng committed
158

Lianmin Zheng's avatar
Lianmin Zheng committed
159
    # HTTP server
Lianmin Zheng's avatar
Lianmin Zheng committed
160
161
    host: str = "127.0.0.1"
    port: int = 30000
Lianmin Zheng's avatar
Lianmin Zheng committed
162
163
    skip_server_warmup: bool = False
    warmups: Optional[str] = None
164
    nccl_port: Optional[int] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
165

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

Lianmin Zheng's avatar
Lianmin Zheng committed
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
    tbo_token_distribution_threshold: float = 0.48
381
    enable_torch_compile: bool = False
382
    torch_compile_max_bs: int = 32
383
    torchao_config: str = ""
384
    enable_nan_detection: bool = False
385
    enable_p2p_check: bool = False
386
    triton_attention_reduce_in_fp32: bool = False
387
    triton_attention_num_kv_splits: int = 8
388
    triton_attention_split_tile_size: Optional[int] = None
389
    num_continuous_decode_steps: int = 1
390
    delete_ckpt_after_loading: bool = False
391
    enable_memory_saver: bool = False
392
    allow_auto_truncate: bool = False
393
    enable_custom_logit_processor: bool = False
394
    flashinfer_mla_disable_ragged: bool = False
395
    disable_shared_experts_fusion: bool = False
396
    disable_chunked_prefix_cache: bool = False
Lianmin Zheng's avatar
Lianmin Zheng committed
397
    disable_fast_image_processor: bool = False
398
    keep_mm_feature_on_device: bool = False
399
    enable_return_hidden_states: bool = False
400
    scheduler_recv_interval: int = 1
401
    numa_node: Optional[List[int]] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
402
    enable_deterministic_inference: bool = False
403

404
405
406
407
408
    # Dynamic batch tokenizer
    enable_dynamic_batch_tokenizer: bool = False
    dynamic_batch_tokenizer_batch_size: int = 32
    dynamic_batch_tokenizer_batch_timeout: float = 0.002

409
410
411
412
    # 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
413
    debug_tensor_dump_prefill_only: bool = False
414

Lianmin Zheng's avatar
Lianmin Zheng committed
415
    # PD disaggregation: can be "null" (not disaggregated), "prefill" (prefill-only), or "decode" (decode-only)
416
    disaggregation_mode: Literal["null", "prefill", "decode"] = "null"
417
    disaggregation_transfer_backend: str = "mooncake"
418
    disaggregation_bootstrap_port: int = 8998
Byron Hsu's avatar
Byron Hsu committed
419
420
421
    disaggregation_decode_tp: Optional[int] = None
    disaggregation_decode_dp: Optional[int] = None
    disaggregation_prefill_pp: Optional[int] = 1
422
    disaggregation_ib_device: Optional[str] = None
423
    disaggregation_decode_enable_offload_kvcache: bool = False
424
    num_reserved_decode_tokens: int = 512  # used for decode kv cache offload in PD
425
426
427
    # FIXME: hack to reduce ITL when decode bs is small
    disaggregation_decode_polling_interval: int = 1

Lianmin Zheng's avatar
Lianmin Zheng committed
428
    # For model weight update and weight loading
429
    custom_weight_loader: Optional[List[str]] = None
430
    weight_loader_disable_mmap: bool = False
431
432
433
434
    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

435
436
437
438
    # For PD-Multiplexing
    enable_pdmux: bool = False
    sm_group_num: int = 3

Lianmin Zheng's avatar
Lianmin Zheng committed
439
440
441
442
443
444
    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
445

Lianmin Zheng's avatar
Lianmin Zheng committed
446
447
448
449
450
451
        # 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)

452
453
        # 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
454
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

        # 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()
504

Lianmin Zheng's avatar
Lianmin Zheng committed
505
506
507
508
509
        # Handle deterministic inference.
        self._handle_deterministic_inference()

        # Handle any other necessary validations.
        self._handle_other_validations()
510

511
    def _handle_deprecated_args(self):
Lianmin Zheng's avatar
Lianmin Zheng committed
512
        pass
513

514
    def _handle_missing_default_values(self):
Lianmin Zheng's avatar
Lianmin Zheng committed
515
516
        if self.tokenizer_path is None:
            self.tokenizer_path = self.model_path
517
518
        if self.served_model_name is None:
            self.served_model_name = self.model_path
519
520
        if self.device is None:
            self.device = get_device()
521
522
523
        if self.random_seed is None:
            self.random_seed = random.randint(0, 1 << 30)

524
525
526
527
528
529
    def _handle_gpu_memory_settings(self, gpu_mem):
        """
        Configure GPU memory-dependent settings including mem_fraction_static,
        chunked_prefill_size, cuda_graph_max_bs, and cuda_graph_bs.
        """
        # Set mem fraction static
Lianmin Zheng's avatar
Lianmin Zheng committed
530
        if self.mem_fraction_static is None:
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
            if gpu_mem is not None:
                # GPU memory capacity = model weights + KV cache pool + activations + cuda graph buffers
                # mem_fraction_static = (model weights + KV cache pool) / GPU memory capacity.

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

                if gpu_mem < 20 * 1024:
                    # T4, 4080. (chunked_prefill_size 2k, cuda_graph_max_bs 8)
                    reserved_mem = (2.8 + parallel_size / 10) * 1024
547
548
                elif gpu_mem < 50 * 1024:
                    # A10, L40, 4090, 5090. (chunked_prefill_size 2k, cuda_graph_max_bs 16 if tp < 4 else 80)
549
550
                    reserved_mem = (2.8 + parallel_size / 10) * 1024
                elif gpu_mem < 90 * 1024:
551
                    # H100, A100. (chunked_prefill_size 8k, cuda_graph_max_bs 256 if tp < 4 else 512)
552
                    reserved_mem = (12 + parallel_size / 2) * 1024
553
554
555
                elif gpu_mem < 100 * 1024:
                    # H20. (chunked_prefill_size 8k, cuda_graph_max_bs 512)
                    reserved_mem = (15 + parallel_size / 2) * 1024
556
                elif gpu_mem < 160 * 1024:
557
558
                    # H200. (chunked_prefill_size 8k, cuda_graph_max_bs 512)
                    reserved_mem = (15 + parallel_size / 2) * 1024
559
                else:
560
561
562
                    # B200, MI300. (chunked_prefill_size 16k, cuda_graph_max_bs 512)
                    reserved_mem = 32 * 1024

563
                # draft model and larger cuda graph buffers
564
                if self.speculative_algorithm is not None:
565
566
567
568
                    if self.speculative_algorithm == "STANDALONE":
                        # Standalone speculative decoding needs more memory than other speculative
                        # decoding algorithms since the draft model is typically larger.
                        reserved_mem += 6 * 1024
569
                    elif self.speculative_algorithm != "NGRAM":
570
                        reserved_mem += 2 * 1024
571
572
573
574
                if self.enable_dp_attention:
                    reserved_mem += 4 * 1024

                self.mem_fraction_static = round((gpu_mem - reserved_mem) / gpu_mem, 3)
575
            else:
576
                self.mem_fraction_static = 0.88
577

578
579
            # Lazy init to avoid circular import
            # Multimodal models need more memory for the image processor
580
581
582
            from sglang.srt.configs.model_config import ModelConfig

            model_config = ModelConfig.from_server_args(self)
Lianmin Zheng's avatar
Lianmin Zheng committed
583
584
            if model_config.is_multimodal:
                self.adjust_mem_fraction_for_vlm(model_config)
585

586
        # Set chunked prefill size, which depends on the gpu memory capacity
587
        if self.chunked_prefill_size is None:
588
            if gpu_mem is not None:
589
                if gpu_mem < 50 * 1024:  # T4, 4080, A10, L40, 4090, 5090
590
                    self.chunked_prefill_size = 2048
591
                elif gpu_mem < 160 * 1024:  # H100, H200, A100, H20
592
                    self.chunked_prefill_size = 8192
593
                else:  # B200, MI300
594
                    self.chunked_prefill_size = 16384
595
            else:
596
                self.chunked_prefill_size = 4096
Lianmin Zheng's avatar
Lianmin Zheng committed
597

598
        # Set cuda graph max batch size and cuda graph batch sizes
599
        if self.cuda_graph_max_bs is None:
600
601
602
            if gpu_mem is not None:
                if gpu_mem < 20 * 1024:
                    # T4, 4080
603
                    self.cuda_graph_max_bs = 8
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
                elif gpu_mem < 50 * 1024:
                    # A10, L40, 4090, 5090
                    # Based on detailed statistics, when serving TP1/TP2 models on lower-end GPUs with HBM<25G, you can either disable cuda graph or set `cuda_graph_max_bs` to a very small value to reduce the memory overhead of creating cuda graphs, with almost no impact on performance.
                    # However, when serving models with TP4 or TP8, we need to enable cuda graph to maintain high performance. In this case, we can set `cuda_graph_max_bs` to 80 (half of the default value 160) to reduce the memory overhead of creating cuda graphs. Looking at the logs
                    # from TP4 serving of qwen2-72b, a value of 80 is sufficient and can reduce the memory overhead of creating cuda graphs on lower-end GPUs compared to the original 160, avoiding OOM issues.
                    if self.tp_size < 4:
                        self.cuda_graph_max_bs = 16
                    else:
                        self.cuda_graph_max_bs = 80
                elif gpu_mem < 90 * 1024:
                    # H100, A100
                    if self.tp_size < 4:
                        self.cuda_graph_max_bs = 256
                    else:
                        self.cuda_graph_max_bs = 512
619
                else:
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
                    # H20, H200, B200, MI300
                    self.cuda_graph_max_bs = 512
            else:
                # Default fallback
                self.cuda_graph_max_bs = 160

        if self.cuda_graph_bs is None:
            self.cuda_graph_bs = self._generate_cuda_graph_batch_sizes()

    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))
                + list(range(512, self.cuda_graph_max_bs + 1))
            )
        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
658

659
    def _handle_hpu_backends(self):
660
661
662
663
        if self.device == "hpu":
            self.attention_backend = "torch_native"
            self.sampling_backend = "pytorch"

664
    def _handle_cpu_backends(self):
665
666
667
668
669
        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
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
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
    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

748
    def _handle_sampling_backend(self):
749
        if self.sampling_backend is None:
750
751
752
753
            self.sampling_backend = (
                "flashinfer" if is_flashinfer_available() else "pytorch"
            )

754
    def _handle_attention_backend_compatibility(self):
755
        if self.attention_backend == "torch_native":
756
            logger.warning(
757
758
759
                "Cuda graph is disabled because of using torch native attention backend"
            )
            self.disable_cuda_graph = True
760

761
762
763
764
765
766
767
768
769
        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"

770
        if is_npu() and self.attention_backend in ["ascend", "hybrid_linear_attn"]:
771
772
773
774
775
            logger.warning(
                "At this moment Ascend attention backend only supports a page_size of 128, change page_size to 128."
            )
            self.page_size = 128

776
777
778
779
        if (
            self.attention_backend == "flashmla"
            or self.decode_attention_backend == "flashmla"
        ):
Lianmin Zheng's avatar
Lianmin Zheng committed
780
781
782
783
784
            logger.warning(
                "FlashMLA only supports a page_size of 64, change page_size to 64."
            )
            self.page_size = 64

785
786
787
788
        if (
            self.attention_backend == "cutlass_mla"
            or self.decode_attention_backend == "cutlass_mla"
        ):
Lianmin Zheng's avatar
Lianmin Zheng committed
789
790
791
792
793
            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
794
795
796
797
        if (
            self.attention_backend == "trtllm_mla"
            or self.decode_attention_backend == "trtllm_mla"
        ):
798
799
800
801
802
803
804
805
806
807
            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
808
809
810
811
812

            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."
                )
813

814
815
816
817
818
        if (
            self.attention_backend == "trtllm_mha"
            or self.decode_attention_backend == "trtllm_mha"
            or self.prefill_attention_backend == "trtllm_mha"
        ):
819
820
821
822
823
824
825
826
827
828
829
            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

830
831
        if self.attention_backend == "dual_chunk_flash_attn":
            logger.warning(
832
                "Mixed chunk, radix cache, and cuda graphs are disabled because of using dual chunk flash attention backend"
833
834
835
836
837
            )
            self.enable_mixed_chunk = False
            self.disable_cuda_graph = True
            self.disable_radix_cache = True

838
    def _handle_page_size(self):
839
840
841
        if self.page_size is None:
            self.page_size = 1

842
    def _handle_amd_specifics(self):
843
844
845
        if is_hip():
            self.triton_attention_num_kv_splits = 16

846
    def _handle_grammar_backend(self):
847
848
        if self.grammar_backend is None:
            self.grammar_backend = "xgrammar"
849

850
    def _handle_data_parallelism(self):
851
852
        if self.dp_size == 1:
            self.enable_dp_attention = False
853
            self.enable_dp_lm_head = False
854

Ke Bao's avatar
Ke Bao committed
855
        if self.enable_dp_attention:
856
            self.schedule_conservativeness = self.schedule_conservativeness * 0.3
Lianmin Zheng's avatar
Lianmin Zheng committed
857
858
            assert self.tp_size % self.dp_size == 0
            self.chunked_prefill_size = self.chunked_prefill_size // self.dp_size
859
            logger.warning(
860
                f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
861
            )
862

863
864
865
        if self.enable_dp_lm_head:
            assert (
                self.enable_dp_attention
866
            ), "Please enable dp attention when setting enable_dp_lm_head. "
867

868
    def _handle_moe_kernel_config(self):
869
        if self.moe_runner_backend == "flashinfer_cutlass":
870
871
872
            assert (
                self.quantization == "modelopt_fp4"
            ), "modelopt_fp4 quantization is required for Flashinfer MOE"
873
874
875
876
            assert self.ep_size in [
                1,
                self.tp_size,
            ], "The expert parallel size must be 1 or the same as the tensor parallel size"
877

878
        if self.moe_runner_backend == "flashinfer_trtllm":
879
880
881
882
883
884
885
            assert (
                self.quantization == "modelopt_fp4" or self.quantization == "fp8"
            ), "modelopt_fp4 quantization is required for Flashinfer TRTLLM MoE"
            self.disable_shared_experts_fusion = True
            logger.warning(
                "FlashInfer TRTLLM MoE is enabled. --disable-shared-experts-fusion is automatically set."
            )
886

887
    def _handle_deepep_moe(self):
888
        if self.moe_a2a_backend == "deepep":
889
890
891
            if self.deepep_mode == "normal":
                logger.warning("Cuda graph is disabled because deepep_mode=`normal`")
                self.disable_cuda_graph = True
892
            self.ep_size = self.tp_size
Lianmin Zheng's avatar
Lianmin Zheng committed
893
            logger.warning(
894
895
                f"DeepEP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
            )
896

897
    def _handle_eplb_and_dispatch(self):
898
899
        if self.enable_eplb and (self.expert_distribution_recorder_mode is None):
            self.expert_distribution_recorder_mode = "stat"
900
            logger.warning(
901
                "EPLB is enabled. The expert_distribution_recorder_mode is automatically set."
902
903
904
905
906
907
908
            )

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

909
        if self.enable_eplb:
910
            assert self.ep_size > 1
911

912
    def _handle_expert_distribution_metrics(self):
913
914
915
916
917
        if self.enable_expert_distribution_metrics and (
            self.expert_distribution_recorder_mode is None
        ):
            self.expert_distribution_recorder_mode = "stat"

918
        if self.expert_distribution_recorder_buffer_size is None:
919
920
            if (x := self.eplb_rebalance_num_iterations) is not None:
                self.expert_distribution_recorder_buffer_size = x
921
922
923
            elif self.expert_distribution_recorder_mode is not None:
                self.expert_distribution_recorder_buffer_size = 1000

924
    def _handle_pipeline_parallelism(self):
Lianmin Zheng's avatar
Lianmin Zheng committed
925
926
927
928
929
930
        if self.pp_size > 1:
            self.disable_overlap_schedule = True
            logger.warning(
                "Pipeline parallelism is incompatible with overlap schedule."
            )

931
    def _handle_hicache(self):
932
        if self.hicache_storage_backend == "mooncake":
933
934
935
936
937
938
939
940
941
            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"
                )
942

943
944
945
946
947
948
949
        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"
                )

950
    def _handle_speculative_decoding(self):
951
952
953
        if self.speculative_algorithm == "NEXTN":
            self.speculative_algorithm = "EAGLE"

954
        if self.speculative_algorithm in ("EAGLE", "EAGLE3", "STANDALONE"):
955
            if self.speculative_algorithm == "STANDALONE" and self.enable_dp_attention:
956
                # TODO: support dp attention for standalone speculative decoding
957
958
959
                raise ValueError(
                    "Currently standalone speculative decoding does not support dp attention."
                )
960
            if self.max_running_requests is None:
961
                self.max_running_requests = 48
962
            self.disable_overlap_schedule = True
Lianmin Zheng's avatar
Lianmin Zheng committed
963
            logger.warning(
964
                "Overlap scheduler is disabled because of using "
965
                "eagle speculative decoding."
966
            )
967
968
969
970
971
972
            if self.enable_mixed_chunk:
                self.enable_mixed_chunk = False
                logger.warning(
                    "Mixed chunked prefill is disabled because of using "
                    "eagle speculative decoding."
                )
973

Lianmin Zheng's avatar
Lianmin Zheng committed
974
            model_arch = self.get_hf_config().architectures[0]
strgrb's avatar
strgrb committed
975
976
977
            if model_arch in [
                "DeepseekV3ForCausalLM",
                "Glm4MoeForCausalLM",
Yuan Luo's avatar
Yuan Luo committed
978
                "BailingMoeForCausalLM",
strgrb's avatar
strgrb committed
979
980
                "BailingMoeV2ForCausalLM",
            ]:
981
982
983
984
985
986
                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."
                    )
987

988
989
990
991
992
993
994
995
996
            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,
997
                ) = auto_choose_speculative_params(self)
998

999
1000
1001
1002
1003
1004
1005
1006
1007
1008
            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."
                    )

1009
1010
1011
1012
            if (
                self.speculative_eagle_topk == 1
                and self.speculative_num_draft_tokens != self.speculative_num_steps + 1
            ):
Lianmin Zheng's avatar
Lianmin Zheng committed
1013
                logger.warning(
1014
1015
1016
                    "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
1017

1018
1019
1020
1021
1022
1023
1024
1025
1026
            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."
                )

1027
        if self.speculative_algorithm == "NGRAM":
1028
1029
            if not self.device.startswith("cuda"):
                raise ValueError(
1030
                    "Ngram speculative decoding only supports CUDA device."
1031
1032
1033
1034
1035
                )
            if self.max_running_requests is None:
                self.max_running_requests = 48
            self.disable_overlap_schedule = True
            self.enable_mixed_chunk = False
1036
            self.speculative_eagle_topk = self.speculative_ngram_max_bfs_breadth
1037
1038
            if self.speculative_num_draft_tokens is None:
                self.speculative_num_draft_tokens = (
1039
                    self.speculative_ngram_max_match_window_size
1040
1041
1042
                )
            logger.warning(
                "The overlap scheduler and mixed chunked prefill are disabled because of "
1043
                "using ngram speculative decoding."
1044
            )
1045

1046
1047
1048
1049
1050
1051
1052
1053
1054
            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."
                )
            if self.enable_dp_attention:
1055
                # TODO: support dp attention for ngram speculative decoding
1056
                raise ValueError(
1057
                    "Currently ngram speculative decoding does not support dp attention."
1058
                )
1059
1060

    def _handle_load_format(self):
1061
1062
1063
1064
1065
        if (
            self.load_format == "auto" or self.load_format == "gguf"
        ) and check_gguf_file(self.model_path):
            self.quantization = self.load_format = "gguf"

1066
1067
        if is_remote_url(self.model_path):
            self.load_format = "remote"
1068

1069
1070
        if self.custom_weight_loader is None:
            self.custom_weight_loader = []
1071

1072
1073
1074
1075
1076
1077
1078
1079
        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"

1080
    def _handle_disaggregation(self):
Byron Hsu's avatar
Byron Hsu committed
1081
1082
1083
1084
1085
1086
1087
1088
        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
1089
            self.disable_radix_cache = True
1090
            logger.warning("KV cache is forced as chunk cache for decode server")
1091
1092
1093
1094
1095
1096
1097

            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
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
        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
1108

1109
    def _handle_tokenizer_batching(self):
1110
1111
1112
1113
1114
1115
        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."
            )

1116
    def _handle_environment_variables(self):
1117
1118
1119
        os.environ["SGLANG_ENABLE_TORCH_COMPILE"] = (
            "1" if self.enable_torch_compile else "0"
        )
Yi Zhang's avatar
Yi Zhang committed
1120
        os.environ["SGLANG_MAMBA_SSM_DTYPE"] = self.mamba_ssm_dtype
1121
1122
1123
        os.environ["SGLANG_DISABLE_OUTLINES_DISK_CACHE"] = (
            "1" if self.disable_outlines_disk_cache else "0"
        )
1124
1125
1126
        os.environ["SGLANG_ENABLE_DETERMINISTIC_INFERENCE"] = (
            "1" if self.enable_deterministic_inference else "0"
        )
1127

1128
    def _handle_cache_compatibility(self):
1129
1130
1131
1132
1133
1134
        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."
            )

1135
1136
1137
1138
1139
1140
1141
1142
        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."
            )

1143
    def _handle_metrics_labels(self):
1144
1145
        if (
            not self.tokenizer_metrics_custom_labels_header
1146
            and self.tokenizer_metrics_allowed_custom_labels
1147
1148
        ):
            raise ValueError(
1149
                "Please set --tokenizer-metrics-custom-labels-header when setting --tokenizer-metrics-allowed-custom-labels."
1150
1151
            )

1152
    def _handle_deterministic_inference(self):
1153
        if self.enable_deterministic_inference:
1154
            # Check sampling backend
1155
1156
1157
1158
            self.sampling_backend = "pytorch"
            logger.warning(
                "Sampling backend is set to pytorch for deterministic inference."
            )
1159
1160
1161
1162
1163
1164
1165

            # 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."
                )

1166
            # Currently, only FA3 supports radix cache. Support for other backends is in progress
1167
1168
1169
            if self.attention_backend != "fa3":
                self.disable_radix_cache = True
                logger.warning(
1170
                    f"Currently radix cache is not compatible with {self.attention_backend} attention backend for deterministic inference. It will be supported in the future."
1171
                )
1172
1173
1174

            # Check TP size
            if self.tp_size > 1:
1175
1176
1177
1178
                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."
1179
1180
                )

1181
    def _handle_other_validations(self):
fzyzcjy's avatar
fzyzcjy committed
1182
        pass
1183

Lianmin Zheng's avatar
Lianmin Zheng committed
1184
1185
    @staticmethod
    def add_cli_args(parser: argparse.ArgumentParser):
Lianmin Zheng's avatar
Lianmin Zheng committed
1186
        # Model and tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
1187
1188
        parser.add_argument(
            "--model-path",
1189
            "--model",
Lianmin Zheng's avatar
Lianmin Zheng committed
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
            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
1200
1201
1202
1203
1204
1205
1206
1207
1208
        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.",
        )
1209
1210
1211
1212
1213
1214
        parser.add_argument(
            "--tokenizer-worker-num",
            type=int,
            default=ServerArgs.tokenizer_worker_num,
            help="The worker num of the tokenizer manager.",
        )
1215
1216
1217
        parser.add_argument(
            "--skip-tokenizer-init",
            action="store_true",
1218
            help="If set, skip init tokenizer and pass input_ids in generate request.",
1219
        )
1220
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
1221
1222
1223
            "--load-format",
            type=str,
            default=ServerArgs.load_format,
1224
            choices=LOAD_FORMAT_CHOICES,
Lianmin Zheng's avatar
Lianmin Zheng committed
1225
1226
1227
1228
1229
1230
1231
1232
1233
            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, '
1234
            "which is mainly for profiling."
1235
1236
            '"gguf" will load the weights in the gguf format. '
            '"bitsandbytes" will load the weights using bitsandbytes '
1237
1238
1239
1240
            "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
1241
        )
1242
1243
1244
1245
1246
1247
1248
        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,
        )
1249
1250
1251
1252
1253
        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
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
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
        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
1325
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
1326
            "--dtype",
Cody Yu's avatar
Cody Yu committed
1327
            type=str,
Lianmin Zheng's avatar
Lianmin Zheng committed
1328
            default=ServerArgs.dtype,
Ying Sheng's avatar
Ying Sheng committed
1329
1330
            choices=["auto", "half", "float16", "bfloat16", "float", "float32"],
            help="Data type for model weights and activations.\n\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
1331
            '* "auto" will use FP16 precision for FP32 and FP16 models, and '
Ying Sheng's avatar
Ying Sheng committed
1332
            "BF16 precision for BF16 models.\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
1333
1334
1335
1336
            '* "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
1337
1338
            '* "float32" for FP32 precision.',
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1339
1340
1341
1342
        parser.add_argument(
            "--quantization",
            type=str,
            default=ServerArgs.quantization,
1343
            choices=QUANTIZATION_CHOICES,
Lianmin Zheng's avatar
Lianmin Zheng committed
1344
1345
            help="The quantization method.",
        )
1346
1347
1348
1349
1350
1351
1352
1353
1354
        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. ",
        )
1355
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
1356
            "--kv-cache-dtype",
1357
            type=str,
Lianmin Zheng's avatar
Lianmin Zheng committed
1358
1359
1360
            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+.',
1361
        )
1362

1363
        # Memory and scheduling
Lianmin Zheng's avatar
Lianmin Zheng committed
1364
1365
1366
1367
        parser.add_argument(
            "--mem-fraction-static",
            type=float,
            default=ServerArgs.mem_fraction_static,
1368
            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
1369
        )
1370
1371
1372
1373
1374
1375
        parser.add_argument(
            "--max-running-requests",
            type=int,
            default=ServerArgs.max_running_requests,
            help="The maximum number of running requests.",
        )
1376
1377
1378
1379
1380
1381
        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.",
        )
1382
1383
1384
1385
        parser.add_argument(
            "--max-total-tokens",
            type=int,
            default=ServerArgs.max_total_tokens,
1386
1387
            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.",
1388
        )
1389
1390
1391
1392
        parser.add_argument(
            "--chunked-prefill-size",
            type=int,
            default=ServerArgs.chunked_prefill_size,
1393
            help="The maximum number of tokens in a chunk for the chunked prefill. Setting this to -1 means disabling chunked prefill.",
1394
1395
1396
1397
1398
1399
1400
        )
        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
1401
        parser.add_argument(
1402
            "--schedule-policy",
Lianmin Zheng's avatar
Lianmin Zheng committed
1403
            type=str,
1404
            default=ServerArgs.schedule_policy,
1405
            choices=["lpm", "random", "fcfs", "dfs-weight", "lof", "priority"],
1406
            help="The scheduling policy of the requests.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1407
        )
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
        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).",
        )
1426
1427
1428
1429
        parser.add_argument(
            "--schedule-conservativeness",
            type=float,
            default=ServerArgs.schedule_conservativeness,
1430
            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.",
1431
        )
1432
1433
1434
1435
1436
1437
        parser.add_argument(
            "--page-size",
            type=int,
            default=ServerArgs.page_size,
            help="The number of tokens in a page.",
        )
tarinkk's avatar
tarinkk committed
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
        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
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
        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.",
        )
1462

Lianmin Zheng's avatar
Lianmin Zheng committed
1463
1464
1465
1466
1467
1468
1469
        # 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
1470
        parser.add_argument(
1471
            "--tensor-parallel-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
1472
            "--tp-size",
Lianmin Zheng's avatar
Lianmin Zheng committed
1473
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
1474
            default=ServerArgs.tp_size,
1475
            help="The tensor parallelism size.",
1476
        )
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
        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.",
        )
1490
1491
1492
        parser.add_argument(
            "--stream-interval",
            type=int,
Lianmin Zheng's avatar
Lianmin Zheng committed
1493
            default=ServerArgs.stream_interval,
1494
            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",
1495
        )
1496
1497
1498
1499
1500
        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
1501
1502
1503
1504
        parser.add_argument(
            "--random-seed",
            type=int,
            default=ServerArgs.random_seed,
1505
            help="The random seed.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1506
        )
1507
1508
1509
1510
        parser.add_argument(
            "--constrained-json-whitespace-pattern",
            type=str,
            default=ServerArgs.constrained_json_whitespace_pattern,
Lianmin Zheng's avatar
Lianmin Zheng committed
1511
            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 ]*",
1512
        )
1513
1514
1515
1516
1517
1518
        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.",
        )
1519
1520
1521
1522
1523
1524
        parser.add_argument(
            "--dist-timeout",
            type=int,
            default=ServerArgs.dist_timeout,
            help="Set timeout for torch.distributed initialization.",
        )
1525
1526
1527
1528
        parser.add_argument(
            "--download-dir",
            type=str,
            default=ServerArgs.download_dir,
1529
            help="Model download directory for huggingface.",
1530
        )
1531
1532
1533
1534
1535
1536
        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.",
        )
1537
1538
1539
1540
1541
1542
        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,...",
        )
1543
1544
1545
1546
1547
        parser.add_argument(
            "--sleep-on-idle",
            action="store_true",
            help="Reduce CPU usage when sglang is idle.",
        )
1548
1549

        # Logging
Lianmin Zheng's avatar
Lianmin Zheng committed
1550
1551
1552
1553
        parser.add_argument(
            "--log-level",
            type=str,
            default=ServerArgs.log_level,
1554
            help="The logging level of all loggers.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1555
        )
1556
        parser.add_argument(
1557
1558
1559
1560
            "--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.",
1561
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1562
        parser.add_argument(
1563
            "--log-requests",
Lianmin Zheng's avatar
Lianmin Zheng committed
1564
            action="store_true",
1565
1566
1567
1568
1569
            help="Log metadata, inputs, outputs of all requests. The verbosity is decided by --log-requests-level",
        )
        parser.add_argument(
            "--log-requests-level",
            type=int,
1570
            default=ServerArgs.log_requests_level,
1571
1572
1573
1574
1575
1576
1577
1578
            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
1579
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1580
1581
1582
        parser.add_argument(
            "--show-time-cost",
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
1583
            help="Show time cost of custom marks.",
Lianmin Zheng's avatar
Lianmin Zheng committed
1584
        )
1585
1586
1587
1588
1589
        parser.add_argument(
            "--enable-metrics",
            action="store_true",
            help="Enable log prometheus metrics.",
        )
1590
1591
1592
1593
1594
1595
1596
        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.",
        )
1597
1598
1599
1600
        parser.add_argument(
            "--tokenizer-metrics-custom-labels-header",
            type=str,
            default=ServerArgs.tokenizer_metrics_custom_labels_header,
1601
            help="Specify the HTTP header for passing custom labels for tokenizer metrics.",
1602
1603
        )
        parser.add_argument(
1604
            "--tokenizer-metrics-allowed-custom-labels",
1605
1606
            type=str,
            nargs="+",
1607
1608
            default=ServerArgs.tokenizer_metrics_allowed_custom_labels,
            help="The custom labels allowed for tokenizer metrics. The labels are specified via a dict in "
1609
            "'--tokenizer-metrics-custom-labels-header' field in HTTP requests, e.g., {'label1': 'value1', 'label2': "
1610
            "'value2'} is allowed if '--tokenizer-metrics-allowed-custom-labels label1 label2' is set.",
1611
        )
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
        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.",
        )
1639
1640
1641
        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 "
1642
1643
            "[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')."
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
        )
        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}",
        )
1659
1660
1661
1662
1663
1664
        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.",
        )
1665
1666
1667
1668
        parser.add_argument(
            "--decode-log-interval",
            type=int,
            default=ServerArgs.decode_log_interval,
1669
            help="The log interval of decode batch.",
1670
        )
1671
1672
1673
1674
1675
1676
        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
1677
1678
1679
1680
1681
        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.",
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
        )
        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
1693
        )
1694

1695
        # API related
Liangsheng Yin's avatar
Liangsheng Yin committed
1696
1697
1698
1699
        parser.add_argument(
            "--api-key",
            type=str,
            default=ServerArgs.api_key,
1700
            help="Set API key of the server. It is also used in the OpenAI API compatible server.",
Liangsheng Yin's avatar
Liangsheng Yin committed
1701
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
1702
1703
1704
1705
1706
1707
        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.",
        )
1708
1709
1710
1711
1712
1713
        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
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
        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.",
        )
1726
        parser.add_argument(
1727
            "--file-storage-path",
1728
            type=str,
1729
            default=ServerArgs.file_storage_path,
1730
1731
            help="The path of the file storage in backend.",
        )
1732
1733
1734
1735
1736
        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
1737
1738
1739
1740
1741
1742
1743
        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())}.",
        )
1744
        tool_call_parser_choices = list(FunctionCallParser.ToolCallParserEnum.keys())
1745
1746
1747
        parser.add_argument(
            "--tool-call-parser",
            type=str,
1748
            choices=tool_call_parser_choices,
1749
            default=ServerArgs.tool_call_parser,
1750
            help=f"Specify the parser for handling tool-call interactions. Options include: {tool_call_parser_choices}.",
1751
        )
1752
1753
1754
1755
1756
1757
        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
1758

1759
1760
        # Data parallelism
        parser.add_argument(
1761
            "--data-parallel-size",
1762
1763
1764
            "--dp-size",
            type=int,
            default=ServerArgs.dp_size,
1765
            help="The data parallelism size.",
1766
1767
1768
1769
1770
        )
        parser.add_argument(
            "--load-balance-method",
            type=str,
            default=ServerArgs.load_balance_method,
1771
            help="The load balancing strategy for data parallelism.",
1772
1773
1774
            choices=[
                "round_robin",
                "shortest_queue",
1775
                "minimum_tokens",
1776
1777
            ],
        )
1778
1779
1780
1781
1782
1783
        parser.add_argument(
            "--load-watch-interval",
            type=float,
            default=ServerArgs.load_watch_interval,
            help="The interval of load watching in seconds.",
        )
1784
1785
1786
1787
1788
1789
        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.",
        )
1790

1791
        # Multi-node distributed serving
1792
        parser.add_argument(
1793
            "--dist-init-addr",
1794
            "--nccl-init-addr",  # For backward compatibility. This will be removed in the future.
1795
            type=str,
1796
            help="The host address for initializing distributed backend (e.g., `192.168.0.2:25000`).",
1797
1798
        )
        parser.add_argument(
Liangsheng Yin's avatar
Liangsheng Yin committed
1799
            "--nnodes", type=int, default=ServerArgs.nnodes, help="The number of nodes."
1800
        )
1801
1802
1803
        parser.add_argument(
            "--node-rank", type=int, default=ServerArgs.node_rank, help="The node rank."
        )
1804

Lianmin Zheng's avatar
Lianmin Zheng committed
1805
1806
1807
1808
1809
1810
1811
        # 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,
        )
1812
1813
1814
1815
1816
        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
1817

1818
        # LoRA
1819
1820
1821
1822
1823
1824
        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.",
        )
1825
1826
1827
1828
1829
1830
1831
1832
1833
        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,
1834
            choices=SUPPORTED_LORA_TARGET_MODULES + [LORA_TARGET_ALL_MODULES],
1835
1836
            nargs="*",
            default=None,
1837
1838
1839
            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.",
1840
        )
1841
1842
1843
1844
1845
1846
        parser.add_argument(
            "--lora-paths",
            type=str,
            nargs="*",
            default=None,
            action=LoRAPathAction,
1847
            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}',
1848
1849
1850
1851
1852
        )
        parser.add_argument(
            "--max-loras-per-batch",
            type=int,
            default=8,
1853
1854
            help="Maximum number of adapters for a running batch, include base-only request.",
        )
1855
1856
1857
1858
1859
1860
        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`.",
        )
1861
1862
1863
        parser.add_argument(
            "--lora-backend",
            type=str,
1864
1865
            choices=LORA_BACKEND_CHOICES,
            default=ServerArgs.lora_backend,
1866
            help="Choose the kernel backend for multi-LoRA serving.",
1867
        )
1868
1869
1870
1871
1872
1873
1874
        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.",
        )
1875
1876

        # Kernel backend
1877
1878
1879
        parser.add_argument(
            "--attention-backend",
            type=str,
1880
            choices=ATTENTION_BACKEND_CHOICES,
1881
1882
1883
            default=ServerArgs.attention_backend,
            help="Choose the kernels for attention layers.",
        )
1884
1885
1886
        parser.add_argument(
            "--prefill-attention-backend",
            type=str,
1887
            choices=ATTENTION_BACKEND_CHOICES,
1888
1889
1890
            default=ServerArgs.prefill_attention_backend,
            help="Choose the kernels for prefill attention layers (have priority over --attention-backend).",
        )
1891
1892
1893
        parser.add_argument(
            "--decode-attention-backend",
            type=str,
1894
            choices=ATTENTION_BACKEND_CHOICES,
1895
1896
1897
            default=ServerArgs.decode_attention_backend,
            help="Choose the kernels for decode attention layers (have priority over --attention-backend).",
        )
1898
1899
1900
1901
1902
1903
1904
        parser.add_argument(
            "--sampling-backend",
            type=str,
            choices=["flashinfer", "pytorch"],
            default=ServerArgs.sampling_backend,
            help="Choose the kernels for sampling layers.",
        )
1905
1906
1907
        parser.add_argument(
            "--grammar-backend",
            type=str,
1908
            choices=GRAMMAR_BACKEND_CHOICES,
1909
            default=ServerArgs.grammar_backend,
Lianmin Zheng's avatar
Lianmin Zheng committed
1910
            help="Choose the backend for grammar-guided decoding.",
1911
        )
1912
1913
1914
        parser.add_argument(
            "--mm-attention-backend",
            type=str,
1915
            choices=["sdpa", "fa3", "triton_attn", "ascend_attn"],
1916
1917
1918
            default=ServerArgs.mm_attention_backend,
            help="Set multimodal attention backend.",
        )
1919

1920
1921
1922
1923
        # Speculative decoding
        parser.add_argument(
            "--speculative-algorithm",
            type=str,
1924
            choices=["EAGLE", "EAGLE3", "NEXTN", "STANDALONE", "NGRAM"],
1925
1926
1927
1928
            help="Speculative algorithm.",
        )
        parser.add_argument(
            "--speculative-draft-model-path",
1929
            "--speculative-draft-model",
1930
1931
1932
            type=str,
            help="The path of the draft model weights. This can be a local folder or a Hugging Face repo ID.",
        )
1933
1934
1935
1936
1937
1938
1939
1940
        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.",
        )
1941
1942
1943
1944
1945
1946
1947
1948
1949
        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,
1950
            help="The number of tokens sampled from the draft model in eagle2 each step.",
1951
1952
            default=ServerArgs.speculative_eagle_topk,
        )
1953
1954
1955
        parser.add_argument(
            "--speculative-num-draft-tokens",
            type=int,
1956
            help="The number of tokens sampled from the draft model in Speculative Decoding.",
1957
1958
            default=ServerArgs.speculative_num_draft_tokens,
        )
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
        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,
        )
1971
1972
1973
1974
1975
1976
        parser.add_argument(
            "--speculative-token-map",
            type=str,
            help="The path of the draft model's small vocab table.",
            default=ServerArgs.speculative_token_map,
        )
1977
        parser.add_argument(
1978
            "--speculative-attention-mode",
1979
1980
            type=str,
            choices=["prefill", "decode"],
1981
1982
            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,
1983
        )
1984
        # Ngram speculative decoding
1985
        parser.add_argument(
1986
            "--speculative-ngram-min-match-window-size",
1987
            type=int,
1988
1989
            default=ServerArgs.speculative_ngram_min_match_window_size,
            help="The minimum window size for pattern matching in ngram speculative decoding.",
1990
1991
        )
        parser.add_argument(
1992
            "--speculative-ngram-max-match-window-size",
1993
            type=int,
1994
1995
            default=ServerArgs.speculative_ngram_max_match_window_size,
            help="The maximum window size for pattern matching in ngram speculative decoding.",
1996
1997
        )
        parser.add_argument(
1998
            "--speculative-ngram-min-bfs-breadth",
1999
            type=int,
2000
2001
            default=ServerArgs.speculative_ngram_min_bfs_breadth,
            help="The minimum breadth for BFS (Breadth-First Search) in ngram speculative decoding.",
2002
2003
        )
        parser.add_argument(
2004
            "--speculative-ngram-max-bfs-breadth",
2005
            type=int,
2006
2007
            default=ServerArgs.speculative_ngram_max_bfs_breadth,
            help="The maximum breadth for BFS (Breadth-First Search) in ngram speculative decoding.",
2008
2009
        )
        parser.add_argument(
2010
            "--speculative-ngram-match-type",
2011
2012
            type=str,
            choices=["BFS", "PROB"],
2013
            default=ServerArgs.speculative_ngram_match_type,
2014
2015
2016
            help="The match type for cache tree.",
        )
        parser.add_argument(
2017
            "--speculative-ngram-branch-length",
2018
            type=int,
2019
2020
            default=ServerArgs.speculative_ngram_branch_length,
            help="The branch length for ngram speculative decoding.",
2021
2022
        )
        parser.add_argument(
2023
            "--speculative-ngram-capacity",
2024
            type=int,
2025
2026
            default=ServerArgs.speculative_ngram_capacity,
            help="The cache capacity for ngram speculative decoding.",
2027
        )
2028
2029
2030
2031
2032

        # Expert parallelism
        parser.add_argument(
            "--expert-parallel-size",
            "--ep-size",
Cheng Wan's avatar
Cheng Wan committed
2033
            "--ep",
2034
2035
2036
2037
2038
            type=int,
            default=ServerArgs.ep_size,
            help="The expert parallelism size.",
        )
        parser.add_argument(
2039
2040
            "--moe-a2a-backend",
            type=str,
2041
            choices=["none", "deepep"],
2042
2043
            default=ServerArgs.moe_a2a_backend,
            help="Choose the backend for MoE A2A.",
2044
        )
2045
        parser.add_argument(
2046
2047
2048
2049
2050
2051
2052
2053
            "--moe-runner-backend",
            type=str,
            choices=[
                "auto",
                "triton",
                "triton_kernel",
                "flashinfer_trtllm",
                "flashinfer_cutlass",
2054
                "flashinfer_mxfp4",
2055
                "flashinfer_cutedsl",
2056
2057
2058
            ],
            default=ServerArgs.moe_runner_backend,
            help="Choose the runner backend for MoE.",
2059
2060
        )
        parser.add_argument(
2061
2062
            "--flashinfer-mxfp4-moe-precision",
            type=str,
2063
            choices=["default", "bf16"],
2064
2065
2066
2067
            default=ServerArgs.flashinfer_mxfp4_moe_precision,
            help="Choose the computation precision of flashinfer mxfp4 moe",
        )
        parser.add_argument(
2068
2069
            "--enable-flashinfer-allreduce-fusion",
            action="store_true",
2070
            help="Enable FlashInfer allreduce fusion with Residual RMSNorm.",
2071
        )
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
        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.",
        )
2120
2121
2122
2123
2124
2125
        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].",
        )
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
        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.",
        )
2155

Yi Zhang's avatar
Yi Zhang committed
2156
2157
2158
2159
2160
        # Mamba Cache
        parser.add_argument(
            "--max-mamba-cache-size",
            type=int,
            default=ServerArgs.max_mamba_cache_size,
2161
            help="The maximum size of the mamba cache.",
Yi Zhang's avatar
Yi Zhang committed
2162
2163
2164
2165
2166
2167
        )
        parser.add_argument(
            "--mamba-ssm-dtype",
            type=str,
            default=ServerArgs.mamba_ssm_dtype,
            choices=["float32", "bfloat16"],
2168
            help="The data type of the SSM states in mamba cache.",
Yi Zhang's avatar
Yi Zhang committed
2169
        )
2170

Lianmin Zheng's avatar
Lianmin Zheng committed
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
        # 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.",
        )
2196
2197
2198
2199
2200
2201
2202
        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
2203
2204
2205
2206
2207
2208
2209
        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",
        )
2210
2211
2212
        parser.add_argument(
            "--hicache-mem-layout",
            type=str,
2213
            choices=["layer_first", "page_first", "page_first_direct"],
2214
2215
2216
            default=ServerArgs.hicache_mem_layout,
            help="The layout of host memory pool for hierarchical cache.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2217
2218
2219
        parser.add_argument(
            "--hicache-storage-backend",
            type=str,
2220
            choices=["file", "mooncake", "hf3fs", "nixl", "aibrix", "dynamic"],
Lianmin Zheng's avatar
Lianmin Zheng committed
2221
            default=ServerArgs.hicache_storage_backend,
2222
2223
2224
2225
            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
2226
        )
pansicheng's avatar
pansicheng committed
2227
2228
2229
2230
2231
2232
2233
        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.",
        )
2234
2235
2236
2237
2238
2239
        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.",
        )
2240
2241
2242
2243
2244
2245
        # 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
2246

2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
        # 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
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
        # 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.",
        )

2316
        # Optimization/debug options
Liangsheng Yin's avatar
Liangsheng Yin committed
2317
        parser.add_argument(
Lianmin Zheng's avatar
Lianmin Zheng committed
2318
            "--disable-radix-cache",
Liangsheng Yin's avatar
Liangsheng Yin committed
2319
            action="store_true",
Ying Sheng's avatar
Ying Sheng committed
2320
            help="Disable RadixAttention for prefix caching.",
Liangsheng Yin's avatar
Liangsheng Yin committed
2321
        )
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
        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.",
        )
2334
2335
2336
        parser.add_argument(
            "--disable-cuda-graph",
            action="store_true",
2337
            help="Disable cuda graph.",
2338
        )
2339
        parser.add_argument(
2340
2341
            "--disable-cuda-graph-padding",
            action="store_true",
2342
            help="Disable cuda graph when padding is needed. Still uses cuda graph when padding is not needed.",
2343
        )
2344
2345
2346
2347
2348
        parser.add_argument(
            "--enable-profile-cuda-graph",
            action="store_true",
            help="Enable profiling of cuda graph capture.",
        )
2349
2350
2351
2352
2353
        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.",
        )
2354
2355
2356
2357
2358
        parser.add_argument(
            "--enable-nccl-nvls",
            action="store_true",
            help="Enable NCCL NVLS for prefill heavy requests when available.",
        )
2359
2360
2361
2362
2363
        parser.add_argument(
            "--enable-symm-mem",
            action="store_true",
            help="Enable NCCL symmetric memory for fast collectives.",
        )
2364
2365
2366
2367
2368
        parser.add_argument(
            "--disable-flashinfer-cutlass-moe-fp4-allgather",
            action="store_true",
            help="Disables quantize before all-gather for flashinfer cutlass moe.",
        )
2369
2370
2371
2372
2373
        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.",
        )
2374
        parser.add_argument(
2375
            "--disable-outlines-disk-cache",
2376
            action="store_true",
2377
            help="Disable disk cache of outlines to avoid possible crashes related to file system or high concurrency.",
2378
        )
2379
2380
2381
2382
2383
        parser.add_argument(
            "--disable-custom-all-reduce",
            action="store_true",
            help="Disable the custom all-reduce kernel and fall back to NCCL.",
        )
2384
2385
2386
2387
2388
        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
2389
        parser.add_argument(
2390
            "--disable-overlap-schedule",
Lianmin Zheng's avatar
Lianmin Zheng committed
2391
            action="store_true",
2392
            help="Disable the overlap scheduler, which overlaps the CPU scheduler with GPU model worker.",
Lianmin Zheng's avatar
Lianmin Zheng committed
2393
        )
2394
2395
2396
        parser.add_argument(
            "--enable-mixed-chunk",
            action="store_true",
2397
            help="Enabling mixing prefill and decode in a batch when using chunked prefill.",
2398
        )
Ke Bao's avatar
Ke Bao committed
2399
2400
2401
        parser.add_argument(
            "--enable-dp-attention",
            action="store_true",
2402
            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
2403
        )
2404
2405
2406
2407
2408
        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.",
        )
2409
2410
2411
2412
2413
        parser.add_argument(
            "--enable-two-batch-overlap",
            action="store_true",
            help="Enabling two micro batches to overlap.",
        )
2414
2415
2416
2417
2418
2419
        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.",
        )
2420
2421
2422
        parser.add_argument(
            "--enable-torch-compile",
            action="store_true",
2423
2424
            help="Optimize the model with torch.compile. Experimental feature.",
        )
2425
        parser.add_argument(
2426
            "--torch-compile-max-bs",
2427
            type=int,
2428
            default=ServerArgs.torch_compile_max_bs,
2429
2430
            help="Set the maximum batch size when using torch compile.",
        )
2431
2432
2433
2434
        parser.add_argument(
            "--torchao-config",
            type=str,
            default=ServerArgs.torchao_config,
2435
            help="Optimize the model with torchao. Experimental feature. Current choices are: int8dq, int8wo, int4wo-<group_size>, fp8wo, fp8dq-per_tensor, fp8dq-per_row",
2436
        )
2437
2438
2439
2440
2441
        parser.add_argument(
            "--enable-nan-detection",
            action="store_true",
            help="Enable the NaN detection for debugging purposes.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2442
        parser.add_argument(
2443
            "--enable-p2p-check",
Lianmin Zheng's avatar
Lianmin Zheng committed
2444
            action="store_true",
2445
            help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
Lianmin Zheng's avatar
Lianmin Zheng committed
2446
        )
2447
        parser.add_argument(
2448
            "--triton-attention-reduce-in-fp32",
2449
            action="store_true",
2450
            help="Cast the intermediate attention results to fp32 to avoid possible crashes related to fp16."
2451
            "This only affects Triton attention kernels.",
2452
        )
2453
2454
2455
2456
2457
2458
        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.",
        )
2459
2460
2461
2462
2463
2464
        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.",
        )
2465
2466
2467
2468
2469
2470
2471
2472
        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.",
        )
2473
2474
2475
2476
2477
        parser.add_argument(
            "--delete-ckpt-after-loading",
            action="store_true",
            help="Delete the model checkpoint after loading the model.",
        )
2478
2479
2480
2481
2482
        parser.add_argument(
            "--enable-memory-saver",
            action="store_true",
            help="Allow saving memory using release_memory_occupation and resume_memory_occupation",
        )
2483
2484
2485
2486
2487
        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.",
        )
2488
2489
2490
2491
2492
        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)",
        )
2493
        parser.add_argument(
2494
            "--flashinfer-mla-disable-ragged",
2495
            action="store_true",
2496
            help="Not using ragged prefill wrapper when running flashinfer mla",
2497
        )
2498
        parser.add_argument(
2499
2500
2501
            "--disable-shared-experts-fusion",
            action="store_true",
            help="Disable shared experts fusion optimization for deepseek v3/r1.",
2502
        )
2503
2504
2505
2506
2507
        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
2508
2509
2510
2511
2512
        parser.add_argument(
            "--disable-fast-image-processor",
            action="store_true",
            help="Adopt base image processor instead of fast image processor.",
        )
2513
2514
2515
2516
2517
        parser.add_argument(
            "--keep-mm-feature-on-device",
            action="store_true",
            help="Keep multimodal feature tensors on device after processing to save D2H copy.",
        )
2518
2519
2520
2521
2522
        parser.add_argument(
            "--enable-return-hidden-states",
            action="store_true",
            help="Enable returning hidden states with responses.",
        )
2523
2524
2525
2526
2527
2528
        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.",
        )
2529
2530
2531
2532
2533
2534
        parser.add_argument(
            "--numa-node",
            type=int,
            nargs="+",
            help="Sets the numa node for the subprocesses. i-th element corresponds to i-th subprocess.",
        )
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554

        # 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.",
        )
2555
2556
2557
2558
2559
        parser.add_argument(
            "--debug-tensor-dump-prefill-only",
            action="store_true",
            help="Only dump the tensors for prefill requests (i.e. batch size > 1).",
        )
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
        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.",
        )
2577

Lianmin Zheng's avatar
Lianmin Zheng committed
2578
        # PD disaggregation
Byron Hsu's avatar
Byron Hsu committed
2579
2580
2581
        parser.add_argument(
            "--disaggregation-mode",
            type=str,
2582
            default=ServerArgs.disaggregation_mode,
Byron Hsu's avatar
Byron Hsu committed
2583
2584
2585
            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',
        )
2586
2587
2588
2589
        parser.add_argument(
            "--disaggregation-transfer-backend",
            type=str,
            default=ServerArgs.disaggregation_transfer_backend,
2590
            choices=DISAGG_TRANSFER_BACKEND_CHOICES,
2591
2592
            help="The backend for disaggregation transfer. Default is mooncake.",
        )
2593
2594
2595
2596
2597
2598
        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
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
        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.",
        )
2617
2618
2619
2620
        parser.add_argument(
            "--disaggregation-ib-device",
            type=str,
            default=ServerArgs.disaggregation_ib_device,
2621
2622
2623
            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.",
2624
        )
2625
2626
2627
2628
2629
        parser.add_argument(
            "--disaggregation-decode-enable-offload-kvcache",
            action="store_true",
            help="Enable async KV cache offloading on decode server (PD mode).",
        )
2630
2631
2632
2633
2634
2635
        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.",
        )
2636
2637
2638
2639
2640
2641
        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
2642
2643

        # Custom weight loader
2644
2645
2646
2647
2648
2649
2650
        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",
        )
2651
2652
2653
2654
2655
        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
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
        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.",
        )
2674
2675

        # For PD-Multiplexing
2676
2677
2678
2679
2680
        parser.add_argument(
            "--enable-pdmux",
            action="store_true",
            help="Enable PD-Multiplexing, PD running on greenctx stream.",
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2681

2682
2683
2684
2685
2686
2687
        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
2688

2689
2690
2691
2692
2693
2694
2695
        # For deterministic inference
        parser.add_argument(
            "--enable-deterministic-inference",
            action="store_true",
            help="Enable deterministic inference mode with batch invariant ops.",
        )

2696
2697
2698
        # Deprecated arguments
        parser.add_argument(
            "--enable-ep-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2699
2700
            action=DeprecatedAction,
            help="NOTE: --enable-ep-moe is deprecated. Please set `--ep-size` to the same value as `--tp-size` instead.",
2701
2702
2703
        )
        parser.add_argument(
            "--enable-deepep-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2704
2705
            action=DeprecatedAction,
            help="NOTE: --enable-deepep-moe is deprecated. Please set `--moe-a2a-backend` to 'deepep' instead.",
2706
        )
2707
2708
        parser.add_argument(
            "--enable-flashinfer-cutlass-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2709
2710
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-cutlass-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_cutlass' instead.",
2711
        )
2712
2713
        parser.add_argument(
            "--enable-flashinfer-cutedsl-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2714
2715
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-cutedsl-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_cutedsl' instead.",
2716
        )
2717
2718
        parser.add_argument(
            "--enable-flashinfer-trtllm-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2719
2720
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-trtllm-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_trtllm' instead.",
2721
2722
2723
        )
        parser.add_argument(
            "--enable-triton-kernel-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2724
2725
            action=DeprecatedAction,
            help="NOTE: --enable-triton-kernel-moe is deprecated. Please set `--moe-runner-backend` to 'triton_kernel' instead.",
2726
        )
2727
2728
        parser.add_argument(
            "--enable-flashinfer-mxfp4-moe",
Lianmin Zheng's avatar
Lianmin Zheng committed
2729
2730
            action=DeprecatedAction,
            help="NOTE: --enable-flashinfer-mxfp4-moe is deprecated. Please set `--moe-runner-backend` to 'flashinfer_mxfp4' instead.",
2731
        )
2732

2733
2734
2735
2736
2737
2738
2739
        # 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
2740
2741
    @classmethod
    def from_cli_args(cls, args: argparse.Namespace):
2742
        args.tp_size = args.tensor_parallel_size
2743
        args.pp_size = args.pipeline_parallel_size
2744
        args.dp_size = args.data_parallel_size
xiaobochen's avatar
xiaobochen committed
2745
        args.ep_size = args.expert_parallel_size
2746

Lianmin Zheng's avatar
Lianmin Zheng committed
2747
2748
2749
2750
        attrs = [attr.name for attr in dataclasses.fields(cls)]
        return cls(**{attr: getattr(args, attr) for attr in attrs})

    def url(self):
2751
        if is_valid_ipv6_address(self.host):
2752
2753
2754
            return f"http://[{self.host}]:{self.port}"
        else:
            return f"http://{self.host}:{self.port}"
Lianmin Zheng's avatar
Lianmin Zheng committed
2755

Lianmin Zheng's avatar
Lianmin Zheng committed
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
    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

2767
    def check_server_args(self):
2768
        # Check parallel size constraints
2769
        assert (
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
            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."

2780
        assert not (
2781
2782
            self.dp_size > 1 and self.nnodes != 1 and not self.enable_dp_attention
        ), "multi-node data parallel is not supported unless dp attention!"
2783

2784
        assert self.base_gpu_id >= 0, "base_gpu_id must be non-negative"
2785
        assert self.gpu_id_step >= 1, "gpu_id_step must be positive"
2786

Lianmin Zheng's avatar
Lianmin Zheng committed
2787
2788
2789
2790
2791
        assert self.moe_dense_tp_size in {
            1,
            None,
        }, "moe_dense_tp_size only support 1 and None currently"

2792
        # Check LoRA
2793
2794
        self.check_lora_server_args()

2795
2796
2797
2798
2799
2800
2801
        # 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
2802
        # Skip validation if chunked prefill is disabled (i.e., size <= 0).
2803
2804
        # Skip validation if disaggregation mode is decode.
        if self.chunked_prefill_size > 0 and self.disaggregation_mode != "decode":
2805
2806
2807
            assert (
                self.chunked_prefill_size % self.page_size == 0
            ), "chunked_prefill_size must be divisible by page_size"
2808

2809
2810
        # Check multi tokenizer
        assert self.tokenizer_worker_num > 0, "Tokenizer worker num must >= 1"
2811
2812
2813
2814
2815
2816
        self.validate_buckets_rule(
            "--prompt-tokens-buckets", self.prompt_tokens_buckets
        )
        self.validate_buckets_rule(
            "--generation-tokens-buckets", self.generation_tokens_buckets
        )
2817

2818
2819
2820
2821
2822
2823
2824
        # 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."

2825
    def check_lora_server_args(self):
2826
        assert self.max_loras_per_batch > 0, "max_loras_per_batch must be positive"
2827

2828
2829
2830
2831
        # 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
2832
                logger.warning(
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
                    "--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
2843
                self.lora_paths = []
2844
                for lora_path in lora_paths:
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
                    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),
2863
                        )
2864
                    else:
2865
2866
2867
                        raise ValueError(
                            f"Invalid type for item in --lora-paths list: {type(lora_path)}. "
                            "Expected a string or a dictionary."
2868
                        )
2869
                    self.lora_paths.append(lora_ref)
2870
            elif isinstance(self.lora_paths, dict):
2871
2872
                self.lora_paths = [
                    LoRARef(lora_name=k, lora_path=v, pinned=False)
2873
                    for k, v in self.lora_paths.items()
2874
                ]
2875
            elif self.lora_paths is None:
2876
                self.lora_paths = []
2877
2878
2879
2880
2881
            else:
                raise ValueError(
                    f"Invalid type for --lora-paths: {type(self.lora_paths)}. "
                    "Expected a list or a dictionary."
                )
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895

            # 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."
2896

2897
2898
2899
2900
2901
2902
            # 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}"
                )
2903
                assert len(self.lora_paths) <= self.max_loaded_loras, (
2904
2905
2906
2907
                    "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)}"
                )

2908
2909
2910
2911
2912
2913
            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
2914
2915
2916
2917
2918
2919
2920
2921
    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}"
        )

2922
2923
2924
2925
2926
2927
2928
2929
2930
    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",
2931
2932
            "custom",
        ], f"Unsupported {arg_name} rule type: '{rule}'. Must be one of: 'tse', 'default', 'custom'"
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954

        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)}"

2955
        elif rule == "custom":
2956
2957
            assert (
                len(buckets_rule) >= 2
2958
            ), f"{arg_name} custom rule requires at least one bucket value: ['custom', value1, ...]"
2959
2960
2961
            try:
                bucket_values = [float(x) for x in buckets_rule[1:]]
            except ValueError:
2962
                assert False, f"{arg_name} custom rule bucket values must be numeric"
2963
2964
            assert len(set(bucket_values)) == len(
                bucket_values
2965
            ), f"{arg_name} custom rule bucket values should not contain duplicates"
2966
2967
            assert all(
                val >= 0 for val in bucket_values
2968
            ), f"{arg_name} custom rule bucket values should be non-negative"
2969

Lianmin Zheng's avatar
Lianmin Zheng committed
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
    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
3008

Lianmin Zheng's avatar
Lianmin Zheng committed
3009
def prepare_server_args(argv: List[str]) -> ServerArgs:
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
    """
    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.
    """
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
    # 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)

3040
3041
    parser = argparse.ArgumentParser()
    ServerArgs.add_cli_args(parser)
Lianmin Zheng's avatar
Lianmin Zheng committed
3042
    raw_args = parser.parse_args(argv)
3043
3044
3045
3046
    server_args = ServerArgs.from_cli_args(raw_args)
    return server_args


3047
3048
3049
ZMQ_TCP_PORT_DELTA = 233


Lianmin Zheng's avatar
Lianmin Zheng committed
3050
3051
@dataclasses.dataclass
class PortArgs:
3052
3053
3054
3055
3056
3057
    # 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
3058

3059
3060
    # The port for nccl initialization (torch.dist)
    nccl_port: int
3061

3062
3063
3064
    # The ipc filename for rpc call between Engine and Scheduler
    rpc_ipc_name: str

3065
3066
3067
    # The ipc filename for Scheduler to send metrics
    metrics_ipc_name: str

3068
3069
3070
    # The ipc filename for Tokenizer and worker tokenizer
    tokenizer_worker_ipc_name: Optional[str]

3071
    @staticmethod
3072
    def init_new(server_args, dp_rank: Optional[int] = None) -> "PortArgs":
3073
        if server_args.nccl_port is None:
Lianmin Zheng's avatar
Lianmin Zheng committed
3074
            nccl_port = server_args.port + random.randint(100, 1000)
3075
            while True:
Lianmin Zheng's avatar
Lianmin Zheng committed
3076
                if is_port_available(nccl_port):
3077
                    break
Lianmin Zheng's avatar
Lianmin Zheng committed
3078
3079
                if nccl_port < 60000:
                    nccl_port += 42
3080
                else:
Lianmin Zheng's avatar
Lianmin Zheng committed
3081
                    nccl_port -= 43
3082
        else:
Lianmin Zheng's avatar
Lianmin Zheng committed
3083
            nccl_port = server_args.nccl_port
3084

3085
3086
3087
3088
3089
3090
        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
3091
                nccl_port=nccl_port,
3092
                rpc_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
3093
                metrics_ipc_name=f"ipc://{tempfile.NamedTemporaryFile(delete=False).name}",
3094
                tokenizer_worker_ipc_name=None,
3095
3096
3097
3098
3099
            )
        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
3100
3101
3102
            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))
3103
3104
            else:
                dist_init_addr = server_args.dist_init_addr.split(":")
Vincent's avatar
Vincent committed
3105

3106
3107
3108
3109
3110
3111
            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
3112
3113
3114
            detokenizer_port = port_base + 1
            rpc_port = port_base + 2
            metrics_ipc_name = port_base + 3
3115
            if dp_rank is None:
3116
                # TokenizerManager to DataParallelController
3117
                scheduler_input_port = port_base + 4
3118
            else:
3119
                scheduler_input_port = port_base + 4 + 1 + dp_rank
3120
3121
3122
3123

            return PortArgs(
                tokenizer_ipc_name=f"tcp://{dist_init_host}:{port_base}",
                scheduler_input_ipc_name=f"tcp://{dist_init_host}:{scheduler_input_port}",
3124
                detokenizer_ipc_name=f"tcp://{dist_init_host}:{detokenizer_port}",
Lianmin Zheng's avatar
Lianmin Zheng committed
3125
                nccl_port=nccl_port,
3126
3127
                rpc_ipc_name=f"tcp://{dist_init_host}:{rpc_port}",
                metrics_ipc_name=f"tcp://{dist_init_host}:{metrics_ipc_name}",
3128
                tokenizer_worker_ipc_name=None,
3129
            )
3130

3131
3132
3133

class LoRAPathAction(argparse.Action):
    def __call__(self, parser, namespace, values, option_string=None):
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
        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)
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159


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)
3160
3161


3162
3163
3164
3165
def print_deprecated_warning(message: str):
    logger.warning(f"\033[33m{message}\033[0m")


3166
def auto_choose_speculative_params(self: ServerArgs):
3167
3168
3169
3170
3171
    """
    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
3172
    hf_config = self.get_hf_config()
3173
    arch = hf_config.architectures[0]
3174
3175
3176
    if self.speculative_algorithm == "STANDALONE":
        # The default value for standalone speculative decoding
        return (3, 1, 4)
3177
3178
3179
    if arch in ["LlamaForCausalLM"]:
        # The default value for llama
        return (5, 4, 8)
3180
3181
3182
3183
    elif arch in [
        "DeepseekV3ForCausalLM",
        "DeepseekV2ForCausalLM",
        "GptOssForCausalLM",
strgrb's avatar
strgrb committed
3184
3185
        "BailingMoeForCausalLM",
        "BailingMoeV2ForCausalLM",
3186
3187
    ]:
        # The default value for deepseek and gpt-oss
3188
        return (3, 1, 4)
3189
3190
3191
3192
3193
    elif arch in ["Grok1ForCausalLM", "Grok1VForCausalLM"]:
        return (5, 4, 8)
    else:
        # The default value for all other models
        return (5, 4, 8)