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, 03:15 UTC
[V1] Chore: cruft removal (#11724)
Robert Shaw committed
Jan 3, 2025 at 18:25 UTC
ad0d567e1cdc77aff435b20bac918bfd0f55db0a
4 files changed
-11
vllm/entrypoints/llm.py
-2
@@ -225,8 +225,6 @@ class LLM:
225
# Logic to switch between engines is done at runtime instead of import
226
# to avoid import order issues
227
self.engine_class = self.get_engine_class()
228
-
229
- # TODO(rob): enable mp by default (issue with fork vs spawn)
228
self.llm_engine = self.engine_class.from_engine_args(
229
engine_args, usage_context=UsageContext.LLM_CLASS)
230
vllm/v1/engine/core_client.py
-2
@@ -94,8 +94,6 @@ class InprocClient(EngineCoreClient):
94
95
* pushes EngineCoreRequest directly into the EngineCore
96
* pulls EngineCoreOutputs by stepping the EngineCore
97
-
98
- TODO: support asyncio-mode for debugging.
97
"""
98
99
def __init__(self, *args, **kwargs):
vllm/v1/engine/llm_engine.py
-4
@@ -42,8 +42,6 @@ class LLMEngine:
42
use_cached_outputs: bool = False,
43
multiprocess_mode: bool = False,
44
) -> None:
45
-
46
- # TODO: Can we avoid this?
45
self.model_config = vllm_config.model_config
46
47
# Tokenizer (+ ensure liveness if running in another process).
@@ -179,8 +177,6 @@ class LLMEngine:
177
178
return request_outputs
179
182
- # TODO(rob): Can we get rid of these?
183
-
180
def get_model_config(self):
181
return self.model_config
182
vllm/v1/engine/processor.py
-3
@@ -49,9 +49,6 @@ class Processor:
49
cache_config.enable_prefix_caching
50
self.mm_hasher = MMHasher()
51
52
- # TODO: run in an ThreadpoolExecutor or BackgroundProcess.
53
- # This ideally should releases the GIL, so we should not block the
54
- # asyncio loop while this is running.
52
def process_inputs(
53
self,
54
request_id: str,