FastDeploy/docs/parameters.md

简体中文

FastDeploy Parameter Documentation

Parameter Description

When using FastDeploy to deploy models (including offline inference and service deployment), the following parameter configurations are involved. Please note that for offline inference, the parameter configurations are the parameter names as shown below; while when starting the service via command line, the separators in the corresponding parameters need to be changed from _ to -, for example max_model_len becomes --max-model-len in command line.

Parameter Name	Type	Description
port	int	Only required for service deployment, HTTP service port number, default: 8000
metrics_port	int	Only required for service deployment, metrics monitoring port number, default: 8001
max_waiting_time	int	Only required for service deployment, maximum wait time for establishing a connection upon service request. Default: -1 (indicates no wait time limit).
max_concurrency	int	Only required for service deployment, the actual number of connections established by the service, default 512
engine_worker_queue_port	int	FastDeploy internal engine communication port, default: 8002
cache_queue_port	int	FastDeploy internal KVCache process communication port, default: 8003
max_model_len	int	Default maximum supported context length for inference, default: 2048
tensor_parallel_size	int	Default tensor parallelism degree for model, default: 1
data_parallel_size	int	Default data parallelism degree for model, default: 1
block_size	int	KVCache management granularity (Token count), recommended default: 64
max_num_seqs	int	Maximum concurrent number in Decode phase, default: 8
mm_processor_kwargs	dict[str]	Multimodal processor parameter configuration, e.g.: {"image_min_pixels": 3136, "video_fps": 2}
tokenizer	str	Tokenizer name or path, defaults to model path
use_warmup	int	Whether to perform warmup at startup, will automatically generate maximum length data for warmup, enabled by default when automatically calculating KV Cache
limit_mm_per_prompt	dict[str]	Limit the amount of multimodal data per prompt, e.g.: {"image": 10, "video": 3}, default: 1 for all
enable_mm	bool	[DEPRECATED] Whether to support multimodal data (for multimodal models only), default: False
quantization	str	Model quantization strategy, when loading BF16 CKPT, specifying wint4 or wint8 supports lossless online 4bit/8bit quantization
gpu_memory_utilization	float	GPU memory utilization, default: 0.9
num_gpu_blocks_override	int	Preallocated KVCache blocks, this parameter can be automatically calculated by FastDeploy based on memory situation, no need for user configuration, default: None
max_num_batched_tokens	int	Maximum batch token count in Prefill phase, default: None (same as max_model_len)
kv_cache_ratio	float	KVCache blocks are divided between Prefill phase and Decode phase according to kv_cache_ratio ratio, default: 0.75
enable_prefix_caching	bool	Whether to enable Prefix Caching, default: False
swap_space	float	When Prefix Caching is enabled, CPU memory size for KVCache swapping, unit: GB, default: None
enable_chunked_prefill	bool	Enable Chunked Prefill, default: False
max_num_partial_prefills	int	When Chunked Prefill is enabled, maximum concurrent number of partial prefill batches, default: 1
max_long_partial_prefills	int	When Chunked Prefill is enabled, maximum number of long requests in concurrent partial prefill batches, default: 1
long_prefill_token_threshold	int	When Chunked Prefill is enabled, requests with token count exceeding this value are considered long requests, default: max_model_len*0.04
static_decode_blocks	int	During inference, each request is forced to allocate corresponding number of blocks from Prefill's KVCache for Decode use, default: 2
reasoning_parser	str	Specify the reasoning parser to extract reasoning content from model output
use_cudagraph	bool	[DEPRECATED] CUDAGraph is enabled by default since version 2.3. It is recommended to read graph_optimization.md carefully before opening.
graph_optimization_config	dict[str]	Can configure parameters related to calculation graph optimization, the default value is'{"use_cudagraph":true, "graph_opt_level":0}'，Detailed description reference graph_optimization.md
disable_custom_all_reduce	bool	Disable Custom all-reduce, default: False
splitwise_role	str	Whether to enable splitwise inference, default value: mixed, supported parameters: ["mixed", "decode", "prefill"]
innode_prefill_ports	str	Internal engine startup ports for prefill instances (only required for single-machine PD separation), default: None
guided_decoding_backend	str	Specify the guided decoding backend to use, supports auto, xgrammar, off, default: off
guided_decoding_disable_any_whitespace	bool	Whether to disable whitespace generation during guided decoding, default: False
speculative_config	dict[str]	Speculative decoding configuration, only supports standard format JSON string, default: None
dynamic_load_weight	int	Whether to enable dynamic weight loading, default: 0
enable_expert_parallel	bool	Whether to enable expert parallel
enable_logprob	bool	Whether to enable return log probabilities of the output tokens or not. If true, returns the log probabilities of each output token returned in the content of message.If logrpob is not used, this parameter can be omitted when starting
logprobs_mode	str	Indicates the content returned in the logprobs. Supported mode: raw_logprobs, processed_logprobs, raw_logits, processed_logits. Raw means the values before applying logit processors, like bad words. Processed means the values after applying such processors.
served_model_name	str	The model name used in the API. If not specified, the model name will be the same as the --model argument
revision	str	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.
chat_template	str	Specify the template used for model concatenation, It supports both string input and file path input. The default value is None. If not specified, the model's default template will be used.
tool_call_parser	str	Specify the function call parser to be used for extracting function call content from the model's output.
tool_parser_plugin	str	Specify the file path of the tool parser to be registered, so as to register parsers that are not in the code repository. The code format within these parsers must adhere to the format used in the code repository.
load_choices	str	By default, the "default" loader is used for weight loading. To load Torch weights or enable weight acceleration, "default_v1" must be used.
max_encoder_cache	int	Maximum number of tokens in the encoder cache (use 0 to disable).
max_processor_cache	int	Maximum number of bytes(in GiB) in the processor cache (use 0 to disable).

1. Relationship between KVCache allocation, num_gpu_blocks_override and block_size?

During FastDeploy inference, GPU memory is occupied by model weights, preallocated KVCache blocks and model computation intermediate activation values. The preallocated KVCache blocks are determined by num_gpu_blocks_override, with block_size (default: 64) as its unit, meaning one block can store KVCache for 64 Tokens.

In actual inference, it's difficult for users to know how to properly configure num_gpu_blocks_override, so FastDeploy uses the following method to automatically derive and configure this value:

• Load the model, after completing model loading, record current memory usage total_memory_after_load and FastDeploy framework memory usage fd_memory_after_load; note the former is actual GPU memory usage (may include other processes), the latter is memory used by FD framework itself;

• According to user-configured max_num_batched_tokens (default: max_model_len), perform fake prefill computation with corresponding length input data, record current maximum FastDeploy framework memory allocation fd_memory_after_prefill, thus model computation intermediate activation values can be considered as fd_memory_after_prefill - fd_memory_after_load;

  • At this point, available GPU memory for KVCache allocation (taking A800 80G as example) is 80GB * gpu_memory_utilization - total_memory_after_load - (fd_memory_after_prefill - fd_memory_after_load)
  • Based on model KVCache precision (e.g. 8bit/16bit), calculate memory size per block, then calculate total allocatable blocks, assign to num_gpu_blocks_override

In service startup logs, we can find Reset block num, the total_block_num:17220, prefill_kvcache_block_num:12915 in log/fastdeploy.log, where total_block_num is the automatically calculated KVCache block count, multiply by block_size to get total cacheable Tokens.

2. Relationship between kv_cache_ratio, block_size and max_num_seqs?

• FastDeploy divides KVCache between Prefill and Decode phases according to kv_cache_ratio. When configuring this parameter, you can use kv_cache_ratio = average input Tokens / (average input + average output Tokens). Typically input is 3x output, so can be configured as 0.75.
• max_num_seqs is the maximum concurrency in Decode phase, generally can be set to maximum 128, but users can also configure based on KVCache situation, e.g. output KVCache Token amount is decode_token_cache = total_block_num * (1 - kv_cache_ratio) * block_size, to prevent extreme OOM situations, can configure max_num_seqs = decode_token_cache / average output Tokens, not exceeding 128.

3. enable_chunked_prefill parameter description

When enable_chunked_prefill is enabled, the service processes long input sequences through dynamic chunking, significantly improving GPU resource utilization. In this mode, the original max_num_batched_tokens parameter no longer constrains the batch token count in prefill phase (limiting single prefill token count), thus introducing max_num_partial_prefills parameter specifically to limit concurrently processed partial batches.

To optimize scheduling priority for short requests, new max_long_partial_prefills and long_prefill_token_threshold parameter combination is added. The former limits the number of long requests in single prefill batch, the latter defines the token threshold for long requests. The system will prioritize batch space for short requests, thereby reducing short request latency in mixed workload scenarios while maintaining stable throughput.