squall/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[Bugfix] Make EngineArgs use named arguments for config construction (#5285)

Browse Source
This commit is contained in:
Michael Goin 2024-06-05 18:16:56 -04:00 committed by GitHub
parent 3d33e372a1
commit 065aff6c16
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 40 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

66
vllm/engine/arg_utils.py
View File

@ -619,34 +619,48 @@ class EngineArgs:
"BitsAndBytes load format and QLoRA adapter only support " "BitsAndBytes load format and QLoRA adapter only support "
f"'bitsandbytes' quantization, but got {self.quantization}") f"'bitsandbytes' quantization, but got {self.quantization}")
device_config = DeviceConfig(self.device) device_config = DeviceConfig(device=self.device)
model_config = ModelConfig( model_config = ModelConfig(
self.model, self.tokenizer, self.tokenizer_mode, model=self.model,
self.trust_remote_code, self.dtype, self.seed, self.revision, tokenizer=self.tokenizer,
self.code_revision, self.rope_scaling, self.tokenizer_revision, tokenizer_mode=self.tokenizer_mode,
self.max_model_len, self.quantization, trust_remote_code=self.trust_remote_code,
self.quantization_param_path, self.enforce_eager, dtype=self.dtype,
self.max_context_len_to_capture, self.max_seq_len_to_capture, seed=self.seed,
self.max_logprobs, self.disable_sliding_window, revision=self.revision,
self.skip_tokenizer_init, self.served_model_name) code_revision=self.code_revision,
cache_config = CacheConfig(self.block_size, rope_scaling=self.rope_scaling,
self.gpu_memory_utilization, tokenizer_revision=self.tokenizer_revision,
self.swap_space, self.kv_cache_dtype, max_model_len=self.max_model_len,
self.num_gpu_blocks_override, quantization=self.quantization,
model_config.get_sliding_window(), quantization_param_path=self.quantization_param_path,
self.enable_prefix_caching) enforce_eager=self.enforce_eager,
max_context_len_to_capture=self.max_context_len_to_capture,
max_seq_len_to_capture=self.max_seq_len_to_capture,
max_logprobs=self.max_logprobs,
disable_sliding_window=self.disable_sliding_window,
skip_tokenizer_init=self.skip_tokenizer_init,
served_model_name=self.served_model_name)
cache_config = CacheConfig(
block_size=self.block_size,
gpu_memory_utilization=self.gpu_memory_utilization,
swap_space=self.swap_space,
cache_dtype=self.kv_cache_dtype,
num_gpu_blocks_override=self.num_gpu_blocks_override,
sliding_window=model_config.get_sliding_window(),
enable_prefix_caching=self.enable_prefix_caching)
parallel_config = ParallelConfig( parallel_config = ParallelConfig(
self.pipeline_parallel_size, pipeline_parallel_size=self.pipeline_parallel_size,
self.tensor_parallel_size, tensor_parallel_size=self.tensor_parallel_size,
self.worker_use_ray, worker_use_ray=self.worker_use_ray,
self.max_parallel_loading_workers, max_parallel_loading_workers=self.max_parallel_loading_workers,
self.disable_custom_all_reduce, disable_custom_all_reduce=self.disable_custom_all_reduce,
TokenizerPoolConfig.create_config( tokenizer_pool_config=TokenizerPoolConfig.create_config(
self.tokenizer_pool_size, self.tokenizer_pool_size,
self.tokenizer_pool_type, self.tokenizer_pool_type,
self.tokenizer_pool_extra_config, self.tokenizer_pool_extra_config,
), ),
self.ray_workers_use_nsight, ray_workers_use_nsight=self.ray_workers_use_nsight,
distributed_executor_backend=self.distributed_executor_backend) distributed_executor_backend=self.distributed_executor_backend)
speculative_config = SpeculativeConfig.maybe_create_spec_config( speculative_config = SpeculativeConfig.maybe_create_spec_config(
@ -665,10 +679,10 @@ class EngineArgs:
) )
scheduler_config = SchedulerConfig( scheduler_config = SchedulerConfig(
self.max_num_batched_tokens, max_num_batched_tokens=self.max_num_batched_tokens,
self.max_num_seqs, max_num_seqs=self.max_num_seqs,
model_config.max_model_len, max_model_len=model_config.max_model_len,
self.use_v2_block_manager, use_v2_block_manager=self.use_v2_block_manager,
num_lookahead_slots=(self.num_lookahead_slots num_lookahead_slots=(self.num_lookahead_slots
if speculative_config is None else if speculative_config is None else
speculative_config.num_lookahead_slots), speculative_config.num_lookahead_slots),