Support Features on Ascend NPU

This section describes the basic functions and features supported by the Ascend NPU.If you encounter issues or have any questions, please open an issue.

If you want to know the meaning and usage of each parameter, click Server Arguments.

Model and tokenizer

Argument	Defaults	Options	A2	A3
--model-path --model	None	Type: str	✅	✅
--tokenizer-path	None	Type: str	✅	✅
--tokenizer-mode	auto	auto, slow	✅	✅
--tokenizer-worker-num	1	Type: int	✅	✅
--skip-tokenizer-init	False	bool flag (set to enable)	✅	✅
--load-format	auto	auto, safetensors	✅	✅
--model-loader- extra-config	Type: str	✅	✅
--trust-remote-code	False	bool flag (set to enable)	✅	✅
--context-length	None	Type: int	✅	✅
--is-embedding	False	bool flag (set to enable)	✅	✅
--enable-multimodal	None	bool flag (set to enable)	✅	✅
--revision	None	Type: str	✅	✅
--model-impl	auto	auto, sglang, transformers	✅	✅

HTTP server

Argument	Defaults	Options	A2	A3
--host	127.0.0.1	Type: str	✅	✅
--port	30000	Type: int	✅	✅
--skip-server-warmup	False	bool flag (set to enable)	✅	✅
--warmups	None	Type: str	✅	✅
--nccl-port	None	Type: int	✅	✅
--fastapi-root-path	None	Type: str	✅	✅
--grpc-mode	False	bool flag (set to enable)	✅	✅

Quantization and data type

Argument	Defaults	Options	A2	A3	Special
--dtype	auto	auto, float16, bfloat16	✅	✅	❌
--quantization	None	modelslim	✅	✅	❌
--quantization-param-path	None	Type: str	❌	❌	✅
--kv-cache-dtype	auto	auto	✅	✅	❌
--enable-fp32-lm-head	False	bool flag (set to enable)	✅	✅	❌
--modelopt-quant	None	Type: str	❌	❌	✅
--modelopt-checkpoint- restore-path	None	Type: str	❌	❌	✅
--modelopt-checkpoint- save-path	None	Type: str	❌	❌	✅
--modelopt-export-path	None	Type: str	❌	❌	✅
--quantize-and-serve	False	bool flag (set to enable)	❌	❌	✅
--rl-quant-profile	None	Type: str	❌	❌	✅

Memory and scheduling

Argument	Defaults	Options	A2	A3
--mem-fraction-static	None	Type: float	✅	✅
--max-running-requests	None	Type: int	✅	✅
--prefill-max-requests	None	Type: int	✅	✅
--max-queued-requests	None	Type: int	✅	✅
--max-total-tokens	None	Type: int	✅	✅
--chunked-prefill-size	None	Type: int	✅	✅
--max-prefill-tokens	16384	Type: int	✅	✅
--schedule-policy	fcfs	lpm, fcfs	✅	✅
--enable-priority- scheduling	False	bool flag (set to enable)	✅	✅
--schedule-low-priority- values-first	False	bool flag (set to enable)	✅	✅
--priority-scheduling- preemption-threshold	10	Type: int	✅	✅
--schedule-conservativeness	1.0	Type: float	✅	✅
--page-size	128	Type: int	✅	✅
--swa-full-tokens-ratio	0.8	Type: float	✅	✅
--disable-hybrid-swa-memory	False	bool flag (set to enable)	✅	✅
--abort-on-priority- when-disabled	False	bool flag (set to enable)	✅	✅
--enable-dynamic-chunking	False	bool flag (set to enable)	✅	✅

Runtime options

Argument	Defaults	Options	A2	A3
--device	None	Type: str	✅	✅
--tensor-parallel-size --tp-size	1	Type: int	✅	✅
--pipeline-parallel-size --pp-size	1	Type: int	✅	✅
--pp-max-micro-batch-size	None	Type: int	✅	✅
--pp-async-batch-depth	None	Type: int	✅	✅
--stream-interval	1	Type: int	✅	✅
--stream-output	False	bool flag (set to enable)	✅	✅
--random-seed	None	Type: int	✅	✅
--constrained-json- whitespace-pattern	None	Type: str	✅	✅
--constrained-json- disable-any-whitespace	False	bool flag (set to enable)	✅	✅
--watchdog-timeout	300	Type: float	✅	✅
--soft-watchdog-timeout	300	Type: float	✅	✅
--dist-timeout	None	Type: int	✅	✅
--base-gpu-id	0	Type: int	✅	✅
--gpu-id-step	1	Type: int	✅	✅
--sleep-on-idle	False	bool flag (set to enable)	✅	✅
--custom-sigquit-handler	None	Optional[Callable]	✅	✅

Logging

Argument	Defaults	Options	A2	A3	Special
--log-level	info	Type: str	✅	✅	❌
--log-level-http	None	Type: str	✅	✅	❌
--log-requests	False	bool flag (set to enable)	✅	✅	❌
--log-requests-level	2	0, 1, 2, 3	✅	✅	❌
--log-requests-format	text	text, json	✅	✅	❌
--crash-dump-folder	None	Type: str	✅	✅	❌
--enable-metrics	False	bool flag (set to enable)	✅	✅	❌
--enable-metrics-for- all-schedulers	False	bool flag (set to enable)	✅	✅	❌
--tokenizer-metrics- custom-labels-header	x-custom-labels	Type: str	✅	✅	❌
--tokenizer-metrics- allowed-custom-labels	None	List[str]	✅	✅	❌
--bucket-time-to- first-token	None	List[float]	✅	✅	❌
--bucket-inter-token- latency	None	List[float]	✅	✅	❌
--bucket-e2e-request- latency	None	List[float]	✅	✅	❌
--collect-tokens- histogram	False	bool flag (set to enable)	✅	✅	❌
--prompt-tokens-buckets	None	List[str]	✅	✅	❌
--generation-tokens-buckets	None	List[str]	✅	✅	❌
--gc-warning-threshold-secs	0.0	Type: float	✅	✅	❌
--decode-log-interval	40	Type: int	✅	✅	❌
--enable-request-time- stats-logging	False	bool flag (set to enable)	✅	✅	❌
--kv-events-config	None	Type: str	❌	❌	✅
--enable-trace	False	bool flag (set to enable)	✅	✅	❌
--oltp-traces-endpoint	localhost:4317	Type: str	✅	✅	❌

RequestMetricsExporter configuration

Argument	Defaults	Options	A2	A3
--export-metrics-to- file	False	bool flag (set to enable)	✅	✅
--export-metrics-to- file-dir	None	Type: str	✅	✅

API related

Argument	Defaults	Options	A2	A3
--api-key	None	Type: str	✅	✅
--admin-api-key	None	Type: str	✅	✅
--served-model-name	None	Type: str	✅	✅
--weight-version	default	Type: str	✅	✅
--chat-template	None	Type: str	✅	✅
--completion-template	None	Type: str	✅	✅
--enable-cache-report	False	bool flag (set to enable)	✅	✅
--reasoning-parser	None	deepseek-r1	✅	✅
--tool-call-parser	None	llama,pythonic	✅	✅
--sampling-defaults	model	openai, model	✅	✅

Data parallelism

Argument	Defaults	Options	A2	A3
--data-parallel-size --dp-size	1	Type: int	✅	✅
--load-balance-method	round_robin	round_robin, total_requests, total_tokens	✅	✅
--prefill-round-robin-balance	False	bool flag (set to enable)	✅	✅

Multi-node distributed serving

Argument	Defaults	Options	A2	A3
--dist-init-addr --nccl-init-addr	None	Type: str	✅	✅
--nnodes	1	Type: int	✅	✅
--node-rank	0	Type: int	✅	✅

Model override args

Argument	Defaults	Options	A2	A3
--json-model-override- args	{}	Type: str	✅	✅
--preferred-sampling- params	None	Type: str	✅	✅

LoRA

Argument	Defaults	Options	A2	A3	Special
--enable-lora	False	Bool flag (set to enable)	✅	✅	❌
--max-lora-rank	None	Type: int	✅	✅	❌
--lora-target-modules	None	all	✅	✅	❌
--lora-paths	None	Type: List[str] / JSON objects	✅	✅	❌
--max-loras-per-batch	8	Type: int	✅	✅	❌
--max-loaded-loras	None	Type: int	✅	✅	❌
--lora-eviction-policy	lru	lru, fifo	✅	✅	❌
--lora-backend	triton	triton	✅	✅	❌
--max-lora-chunk-size	16	16, 32, 64, 128	❌	❌	✅

Kernel Backends (Attention, Sampling, Grammar, GEMM)

Argument	Defaults	Options	A2	A3	Special
--attention-backend	None	ascend	✅	✅	❌
--prefill-attention-backend	None	ascend	✅	✅	❌
--decode-attention-backend	None	ascend	✅	✅	❌
--sampling-backend	None	pytorch, ascend	✅	✅	❌
--grammar-backend	None	xgrammar	✅	✅	❌
--mm-attention-backend	None	ascend_attn	✅	✅	❌
--nsa-prefill-backend	flashmla_sparse	flashmla_sparse, flashmla_decode, fa3, tilelang, aiter	❌	❌	✅
--nsa-decode-backend	fa3	flashmla_prefill, flashmla_kv, fa3, tilelang, aiter	❌	❌	✅
--fp8-gemm-backend	auto	auto, deep_gemm, flashinfer_trtllm, cutlass, triton, aiter	❌	❌	✅
--disable-flashinfer- autotune	False	bool flag (set to enable)	❌	❌	✅

Speculative decoding

Argument	Defaults	Options	A2	A3	Special
--speculative-algorithm	None	EAGLE3, NEXTN	✅	✅	❌
--speculative-draft-model-path --speculative-draft-model	None	Type: str	✅	✅	❌
--speculative-draft-model- revision	None	Type: str	✅	✅	❌
--speculative-draft-load-format	None	auto	✅	✅	❌
--speculative-num-steps	None	Type: int	✅	✅	❌
--speculative-eagle-topk	None	Type: int	✅	✅	❌
--speculative-num-draft-tokens	None	Type: int	✅	✅	❌
--speculative-accept- threshold-single	1.0	Type: float	❌	❌	✅
--speculative-accept- threshold-acc	1.0	Type: float	❌	❌	✅
--speculative-token-map	None	Type: str	✅	✅	❌
--speculative-attention- mode	prefill	prefill, decode	✅	✅	❌
--speculative-moe-runner- backend	None	auto	✅	✅	❌
--speculative-moe-a2a- backend	None	ascend_fuseep	✅	✅	❌
--speculative-draft-attention-backend	None	ascend	✅	✅	❌
--speculative-draft-model-quantization	None	unquant	✅	✅	❌

Ngram speculative decoding

Argument	Defaults	Options	A2	A3	Experimental
--speculative-ngram- min-match-window-size	1	Type: int	❌	❌	✅
--speculative-ngram- max-match-window-size	12	Type: int	❌	❌	✅
--speculative-ngram- min-bfs-breadth	1	Type: int	❌	❌	✅
--speculative-ngram- max-bfs-breadth	10	Type: int	❌	❌	✅
--speculative-ngram- match-type	BFS	BFS, PROB	❌	❌	✅
--speculative-ngram- branch-length	18	Type: int	❌	❌	✅
--speculative-ngram- capacity	10000000	Type: int	❌	❌	✅

Expert parallelism

Argument	Defaults	Options	A2	A3	Special
--expert-parallel-size --ep-size --ep	1	Type: int	✅	✅	❌
--moe-a2a-backend	none	none, deepep, ascend_fuseep	✅	✅	❌
--moe-runner-backend	auto	auto, triton	✅	✅	❌
--flashinfer-mxfp4- moe-precision	default	default, bf16	❌	❌	✅
--enable-flashinfer- allreduce-fusion	False	bool flag (set to enable)	❌	❌	✅
--deepep-mode	auto	normal, low_latency, auto	✅	✅	❌
--deepep-config	None	Type: str	❌	❌	✅
--ep-num-redundant-experts	0	Type: int	✅	✅	❌
--ep-dispatch-algorithm	None	Type: str	✅	✅	❌
--init-expert-location	trivial	Type: str	✅	✅	❌
--enable-eplb	False	bool flag (set to enable)	✅	✅	❌
--eplb-algorithm	auto	Type: str	✅	✅	❌
--eplb-rebalance-layers- per-chunk	None	Type: int	✅	✅	❌
--eplb-min-rebalancing- utilization-threshold	1.0	Type: float	✅	✅	❌
--expert-distribution- recorder-mode	None	Type: str	✅	✅	❌
--expert-distribution- recorder-buffer-size	None	Type: int	✅	✅	❌
--enable-expert-distribution- metrics	False	bool flag (set to enable)	✅	✅	❌
--moe-dense-tp-size	None	Type: int	✅	✅	❌
--elastic-ep-backend	None	none, mooncake	❌	❌	✅
--mooncake-ib-device	None	Type: str	❌	❌	✅

Mamba Cache

Argument	Defaults	Options	A2	A3
--max-mamba-cache-size	None	Type: int	✅	✅
--mamba-ssm-dtype	float32	float32, bfloat16	✅	✅
--mamba-full-memory-ratio	0.2	Type: float	✅	✅
--mamba-scheduler-strategy	auto	auto, no_buffer, extra_buffer	✅	✅
--mamba-track-interval	256	Type: int	✅	✅

Hierarchical cache

Argument	Defaults	Options	A2	A3	Special
--enable-hierarchical- cache	False	bool flag (set to enable)	✅	✅	❌
--hicache-ratio	2.0	Type: float	✅	✅	❌
--hicache-size	0	Type: int	✅	✅	❌
--hicache-write-policy	write_through	write_back, write_through, write_through_selective	✅	✅	❌
--radix-eviction-policy	lru	lru, lfu	✅	✅	❌
--hicache-io-backend	kernel	kernel_ascend, direct	✅	✅	❌
--hicache-mem-layout	layer_first	page_first_direct, page_first_kv_split	✅	✅	❌
--hicache-storage- backend	None	file	✅	✅	❌
--hicache-storage- prefetch-policy	best_effort	best_effort, wait_complete, timeout	❌	❌	✅
--hicache-storage- backend-extra-config	None	Type: str	❌	❌	✅

LMCache

Argument	Defaults	Options	A2	A3	Special
--enable-lmcache	False	bool flag (set to enable)	❌	❌	✅

Offloading

Argument	Defaults	Options	A2	A3
--cpu-offload-gb	0	Type: int	✅	✅
--offload-group-size	-1	Type: int	✅	✅
--offload-num-in-group	1	Type: int	✅	✅
--offload-prefetch-step	1	Type: int	✅	✅
--offload-mode	cpu	Type: str	✅	✅

Args for multi-item scoring

Argument	Defaults	Options	A2	A3
--multi-item-scoring-delimiter	None	Type: int	✅	✅

Optimization/debug options

Argument	Defaults	Options	A2	A3	Special	Planned
--disable-radix-cache	False	bool flag (set to enable)	✅	✅	❌	❌
--cuda-graph-max-bs	None	Type: int	✅	✅	❌	❌
--cuda-graph-bs	None	List[int]	✅	✅	❌	❌
--disable-cuda-graph	False	bool flag (set to enable)	✅	✅	❌	❌
--disable-cuda-graph- padding	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-profile- cuda-graph	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-cudagraph-gc	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-nccl-nvls	False	bool flag (set to enable)	❌	❌	✅	❌
--enable-symm-mem	False	bool flag (set to enable)	❌	❌	✅	❌
--disable-flashinfer- cutlass-moe-fp4-allgather	False	bool flag (set to enable)	❌	❌	✅	❌
--enable-tokenizer- batch-encode	False	bool flag (set to enable)	✅	✅	❌	❌
--disable-tokenizer- batch-encode	False	bool flag (set to enable)	✅	✅	❌	❌
--disable-outlines- disk-cache	False	bool flag (set to enable)	✅	✅	❌	❌
--disable-custom- all-reduce	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-mscclpp	False	bool flag (set to enable)	❌	❌	✅	❌
--enable-torch- symm-mem	False	bool flag (set to enable)	❌	❌	✅	❌
--disable-overlap -schedule	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-mixed- chunk	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-dp-attention	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-dp-lm-head	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-two- batch-overlap	False	bool flag (set to enable)	❌	❌	❌	✅
--enable-single- batch-overlap	False	bool flag (set to enable)	✅	✅	❌	❌
--tbo-token- distribution-threshold	0.48	Type: float	❌	❌	❌	✅
--enable-torch- compile	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-torch- compile-debug-mode	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-piecewise- cuda-graph	False	bool flag (set to enable)	✅	✅	❌	❌
--piecewise-cuda- graph-tokens	None	Type: JSON list	✅	✅	❌	❌
--piecewise-cuda- graph-compiler	eager	[“eager”, “inductor”]	✅	✅	❌	❌
--torch-compile-max-bs	32	Type: int	✅	✅	❌	❌
--piecewise-cuda- graph-max-tokens	4096	Type: int	✅	✅	❌	❌
--torchao-config	“	Type: str	❌	❌	✅	❌
--enable-nan-detection	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-p2p-check	False	bool flag (set to enable)	❌	❌	✅	❌
--triton-attention- reduce-in-fp32	False	bool flag (set to enable)	❌	❌	✅	❌
--triton-attention- num-kv-splits	8	Type: int	❌	❌	✅	❌
--triton-attention- split-tile-size	None	Type: int	❌	❌	✅	❌
--delete-ckpt- after-loading	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-memory-saver	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-weights- cpu-backup	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-draft-weights- cpu-backup	False	bool flag (set to enable)	✅	✅	❌	❌
--allow-auto-truncate	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-custom- logit-processor	False	bool flag (set to enable)	✅	✅	❌	❌
--flashinfer-mla- disable-ragged	False	bool flag (set to enable)	❌	❌	✅	❌
--disable-shared- experts-fusion	False	bool flag (set to enable)	✅	✅	❌	❌
--disable-chunked- prefix-cache	False	bool flag (set to enable)	✅	✅	❌	❌
--disable-fast- image-processor	False	bool flag (set to enable)	✅	✅	❌	❌
--keep-mm-feature- on-device	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-return- hidden-states	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-return- routed-experts	False	bool flag (set to enable)	✅	✅	❌	❌
--scheduler-recv- interval	1	Type: int	✅	✅	❌	❌
--numa-node	None	List[int]	✅	✅	❌	❌
--rl-on-policy-target	None	fsdp	❌	❌	❌	✅
--enable-layerwise- nvtx-marker	False	bool flag (set to enable)	❌	❌	✅	❌
--enable-attn-tp- input-scattered	False	bool flag (set to enable)	❌	❌	❌	❌
--enable-nsa-prefill- context-parallel	False	bool flag (set to enable)	✅	✅	❌	❌
--enable-fused-qk- norm-rope	False	bool flag (set to enable)	❌	❌	✅	❌

Dynamic batch tokenizer

Argument	Defaults	Options	A2	A3
--enable-dynamic- batch-tokenizer	False	bool flag (set to enable)	✅	✅
--dynamic-batch- tokenizer-batch-size	32	Type: int	✅	✅
--dynamic-batch- tokenizer-batch-timeout	0.002	Type: float	✅	✅

Debug tensor dumps

Argument	Defaults	Options	A2	A3
--debug-tensor-dump- output-folder	None	Type: str	✅	✅
--debug-tensor-dump- layers	None	List[int]	✅	✅
--debug-tensor-dump- input-file	None	Type: str	✅	✅

PD disaggregation

Argument	Defaults	Options	A2	A3	Special
--disaggregation-mode	null	null, prefill, decode	✅	✅	❌
--disaggregation-transfer-backend	mooncake	ascend	✅	✅	❌
--disaggregation-bootstrap-port	8998	Type: int	✅	✅	❌
--disaggregation-decode-tp	None	Type: int	✅	✅	❌
--disaggregation-decode-dp	None	Type: int	✅	✅	❌
--disaggregation-ib-device	None	Type: str	❌	❌	✅
--disaggregation-decode- enable-offload-kvcache	False	bool flag (set to enable)	✅	✅	❌
--disaggregation-decode- enable-fake-auto	False	bool flag (set to enable)	✅	✅	❌
--num-reserved-decode-tokens	512	Type: int	✅	✅	❌
--disaggregation-decode- polling-interval	1	Type: int	✅	✅	❌

Encode prefill disaggregation

Argument	Defaults	Options	A2	A3
--encoder-only	False	bool flag (set to enable)	✅	✅
--language-only	False	bool flag (set to enable)	✅	✅
--encoder-transfer-backend	zmq_to_scheduler	zmq_to_scheduler, zmq_to_tokenizer, mooncake	✅	✅
--encoder-urls	[]	List[str]	✅	✅

Custom weight loader

Argument	Defaults	Options	A2	A3	Special
--custom-weight-loader	None	List[str]	✅	✅	❌
--weight-loader-disable- mmap	False	bool flag (set to enable)	✅	✅	❌
--remote-instance-weight- loader-seed-instance-ip	None	Type: str	✅	✅	❌
--remote-instance-weight- loader-seed-instance-service-port	None	Type: int	✅	✅	❌
--remote-instance-weight- loader-send-weights-group-ports	None	Type: JSON list	✅	✅	❌
--remote-instance-weight- loader-backend	nccl	transfer_engine, nccl	✅	✅	❌
--remote-instance-weight- loader-start-seed-via-transfer-engine	False	bool flag (set to enable)	❌	❌	✅

For PD-Multiplexing

Argument	Defaults	Options	A2	A3	Special
--enable-pdmux	False	bool flag (set to enable)	❌	❌	✅
--pdmux-config-path	None	Type: str	❌	❌	✅
--sm-group-num	8	Type: int	❌	❌	✅

For Multi-Modal

Argument	Defaults	Options	A2	A3
--mm-max-concurrent-calls	32	Type: int	✅	✅
--mm-per-request-timeout	10.0	Type: float	✅	✅
--enable-broadcast-mm- inputs-process	False	bool flag (set to enable)	✅	✅
--mm-process-config	None	Type: JSON / Dict	✅	✅
--mm-enable-dp-encoder	False	bool flag (set to enable)	✅	✅
--limit-mm-data-per-request	None	Type: JSON / Dict	✅	✅

For checkpoint decryption

Argument	Defaults	Options	A2	A3
--decrypted-config-file	None	Type: str	✅	✅
--decrypted-draft-config-file	None	Type: str	✅	✅
--enable-prefix-mm-cache	False	bool flag (set to enable)	✅	✅

For deterministic inference

Argument	Defaults	Options	A2	A3	Planned
--enable-deterministic- inference	False	bool flag (set to enable)	❌	❌	✅

For registering hooks

Argument	Defaults	Options	A2	A3
--forward-hooks	None	Type: JSON list	✅	✅

Configuration file support

Argument	Defaults	Options	A2	A3
--config	None	Type: str	✅	✅

Other Params

The following parameters are not supported because the third-party components that depend on are not compatible with the NPU, like Ktransformer, checkpoint-engine etc.

Argument	Defaults	Options
--checkpoint-engine- wait-weights- before-ready	False	bool flag (set to enable)
--kt-weight-path	None	Type: str
--kt-method	AMXINT4	Type: str
--kt-cpuinfer	None	Type: int
--kt-threadpool-count	2	Type: int
--kt-num-gpu-experts	None	Type: int
--kt-max-deferred- experts-per-token	None	Type: int

The following parameters have some functional deficiencies on community

Argument	Defaults	Options
--enable-double-sparsity	False	bool flag (set to enable)
--ds-channel-config-path	None	Type: str
--ds-heavy-channel-num	32	Type: int
--ds-heavy-token-num	256	Type: int
--ds-heavy-channel-type	qk	Type: str
--ds-sparse-decode- threshold	4096	Type: int
--tool-server	None	Type: str