A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 05:45 UTC
[misc] clean up and unify logging (#10999)
Signed-off-by: youkaichao <youkaichao@gmail.com>
youkaichao committed
Dec 8, 2024 at 17:28 UTC
46004e83a2e0b908f28099d93171bfb4934e4722
2 files changed
+37
-90
vllm/config.py
+34
-39
@@ -2579,45 +2579,40 @@ class VllmConfig:
2579
self.instance_id = random_uuid()[:5]
2580
2581
def __str__(self):
2582
- return ("model=%r, speculative_config=%r, tokenizer=%r, "
2583
- "skip_tokenizer_init=%s, tokenizer_mode=%s, revision=%s, "
2584
- "override_neuron_config=%s, tokenizer_revision=%s, "
2585
- "trust_remote_code=%s, dtype=%s, max_seq_len=%d, "
2586
- "download_dir=%r, load_format=%s, tensor_parallel_size=%d, "
2587
- "pipeline_parallel_size=%d, "
2588
- "disable_custom_all_reduce=%s, quantization=%s, "
2589
- "enforce_eager=%s, kv_cache_dtype=%s, "
2590
- "quantization_param_path=%s, device_config=%s, "
2591
- "decoding_config=%r, observability_config=%r, "
2592
- "seed=%d, served_model_name=%s, "
2593
- "num_scheduler_steps=%d, enable_prefix_caching=%s, "
2594
- "use_async_output_proc=%s, mm_processor_kwargs=%s") % \
2595
- (self.model_config.model, self.speculative_config,
2596
- self.model_config.tokenizer,
2597
- self.model_config.skip_tokenizer_init,
2598
- self.model_config.tokenizer_mode,
2599
- self.model_config.revision,
2600
- self.model_config.override_neuron_config,
2601
- self.model_config.tokenizer_revision,
2602
- self.model_config.trust_remote_code,
2603
- self.model_config.dtype,
2604
- self.model_config.max_model_len,
2605
- self.load_config.download_dir,
2606
- self.load_config.load_format,
2607
- self.parallel_config.tensor_parallel_size,
2608
- self.parallel_config.pipeline_parallel_size,
2609
- self.parallel_config.disable_custom_all_reduce,
2610
- self.model_config.quantization,
2611
- self.model_config.enforce_eager,
2612
- self.cache_config.cache_dtype,
2613
- self.model_config.quantization_param_path,
2614
- self.device_config.device, self.decoding_config,
2615
- self.observability_config, self.model_config.seed,
2616
- self.model_config.served_model_name,
2617
- self.scheduler_config.num_scheduler_steps,
2618
- self.cache_config.enable_prefix_caching,
2619
- self.model_config.use_async_output_proc,
2620
- self.model_config.mm_processor_kwargs)
2582
+ return (
2583
+ f"model={self.model_config.model!r},"
2584
+ f" speculative_config={self.speculative_config!r},"
2585
+ f" tokenizer={self.model_config.tokenizer!r}, "
2586
+ f"skip_tokenizer_init={self.model_config.skip_tokenizer_init},"
2587
+ f" tokenizer_mode={self.model_config.tokenizer_mode}, "
2588
+ f"revision={self.model_config.revision}, "
2589
+ f"override_neuron_config={self.model_config.override_neuron_config},"
2590
+ f" tokenizer_revision={self.model_config.tokenizer_revision}, "
2591
+ f"trust_remote_code={self.model_config.trust_remote_code}, "
2592
+ f"dtype={self.model_config.dtype}, "
2593
+ f"max_seq_len={self.model_config.max_model_len},"
2594
+ f" download_dir={self.load_config.download_dir!r}, "
2595
+ f"load_format={self.load_config.load_format}, "
2596
+ f"tensor_parallel_size={self.parallel_config.tensor_parallel_size},"
2597
+ f" pipeline_parallel_size={self.parallel_config.pipeline_parallel_size}, " # noqa
2598
+ f"disable_custom_all_reduce={self.parallel_config.disable_custom_all_reduce}, " # noqa
2599
+ f"quantization={self.model_config.quantization}, "
2600
+ f"enforce_eager={self.model_config.enforce_eager}, "
2601
+ f"kv_cache_dtype={self.cache_config.cache_dtype}, "
2602
+ f"quantization_param_path={self.model_config.quantization_param_path},"
2603
+ f" device_config={self.device_config.device}, "
2604
+ f"decoding_config={self.decoding_config!r}, "
2605
+ f"observability_config={self.observability_config!r}, "
2606
+ f"seed={self.model_config.seed}, "
2607
+ f"served_model_name={self.model_config.served_model_name}, "
2608
+ f"num_scheduler_steps={self.scheduler_config.num_scheduler_steps}, "
2609
+ f"multi_step_stream_outputs={self.scheduler_config.multi_step_stream_outputs}, " # noqa
2610
+ f"enable_prefix_caching={self.cache_config.enable_prefix_caching}, "
2611
+ f"chunked_prefill_enabled={self.scheduler_config.chunked_prefill_enabled}, " # noqa
2612
+ f"use_async_output_proc={self.model_config.use_async_output_proc}, "
2613
+ f"mm_processor_kwargs={self.model_config.mm_processor_kwargs}, "
2614
+ f"pooler_config={self.model_config.pooler_config!r},"
2615
+ f" compilation_config={self.compilation_config!r}")
2616
2617
2618
_current_vllm_config: Optional[VllmConfig] = None
vllm/engine/llm_engine.py
+3
-51
@@ -247,60 +247,12 @@ class LLMEngine:
247
)
248
249
logger.info(
250
- "Initializing an LLM engine (v%s) with config: "
251
- "model=%r, speculative_config=%r, tokenizer=%r, "
252
- "skip_tokenizer_init=%s, tokenizer_mode=%s, revision=%s, "
253
- "override_neuron_config=%s, tokenizer_revision=%s, "
254
- "trust_remote_code=%s, dtype=%s, max_seq_len=%d, "
255
- "download_dir=%r, load_format=%s, tensor_parallel_size=%d, "
256
- "pipeline_parallel_size=%d, "
257
- "disable_custom_all_reduce=%s, quantization=%s, "
258
- "enforce_eager=%s, kv_cache_dtype=%s, "
259
- "quantization_param_path=%s, device_config=%s, "
260
- "decoding_config=%r, observability_config=%r, "
261
- "seed=%d, served_model_name=%s, "
262
- "num_scheduler_steps=%d, chunked_prefill_enabled=%s "
263
- "multi_step_stream_outputs=%s, enable_prefix_caching=%s, "
264
- "use_async_output_proc=%s, use_cached_outputs=%s, "
265
- "mm_processor_kwargs=%s, pooler_config=%r,"
266
- "compilation_config=%r",
250
+ "Initializing an LLM engine (v%s) with config: %r,"
251
+ "use_cached_outputs=%s, ",
252
VLLM_VERSION,
268
- self.model_config.model,
269
- self.speculative_config,
270
- self.model_config.tokenizer,
271
- self.model_config.skip_tokenizer_init,
272
- self.model_config.tokenizer_mode,
273
- self.model_config.revision,
274
- self.model_config.override_neuron_config,
275
- self.model_config.tokenizer_revision,
276
- self.model_config.trust_remote_code,
277
- self.model_config.dtype,
278
- self.model_config.max_model_len,
279
- self.load_config.download_dir,
280
- self.load_config.load_format,
281
- self.parallel_config.tensor_parallel_size,
282
- self.parallel_config.pipeline_parallel_size,
283
- self.parallel_config.disable_custom_all_reduce,
284
- self.model_config.quantization,
285
- self.model_config.enforce_eager,
286
- self.cache_config.cache_dtype,
287
- self.model_config.quantization_param_path,
288
- self.device_config.device,
289
- self.decoding_config,
290
- self.observability_config,
291
- self.model_config.seed,
292
- self.model_config.served_model_name,
293
- self.scheduler_config.num_scheduler_steps,
294
- self.scheduler_config.chunked_prefill_enabled,
295
- self.scheduler_config.multi_step_stream_outputs,
296
- self.cache_config.enable_prefix_caching,
297
- self.model_config.use_async_output_proc,
253
+ vllm_config,
254
use_cached_outputs,
299
- self.model_config.mm_processor_kwargs,
300
- self.model_config.pooler_config,
301
- vllm_config.compilation_config,
255
)
303
- # TODO(woosuk): Print more configs in debug mode.
256
257
self.log_stats = log_stats
258
self.use_cached_outputs = use_cached_outputs