[Misc] [Core] Implement RFC "Augment BaseExecutor interfaces to enable hardware-agnostic speculative decoding" (#3837)

This commit is contained in:
Cade Daniel
2024-04-09 11:44:15 -07:00
committed by GitHub
parent 6d592eb430
commit e7c7067b45
20 changed files with 451 additions and 275 deletions
+22
View File
@@ -127,6 +127,8 @@ class LLMEngine:
speculative_config=speculative_config,
)
self._initialize_kv_caches()
# If usage stat is enabled, collect relevant info.
if is_usage_stats_enabled():
from vllm.model_executor.model_loader import (
@@ -178,6 +180,26 @@ class LLMEngine:
labels=dict(model_name=model_config.model))
self.stat_logger.info("cache_config", self.cache_config)
def _initialize_kv_caches(self) -> None:
"""Initialize the KV cache in the worker(s).
The workers will determine the number of blocks in both the GPU cache
and the swap CPU cache.
"""
num_gpu_blocks, num_cpu_blocks = (
self.model_executor.determine_num_available_blocks())
if self.cache_config.num_gpu_blocks_override is not None:
num_gpu_blocks_override = self.cache_config.num_gpu_blocks_override
logger.info(f"Overriding {num_gpu_blocks=} with "
f"{num_gpu_blocks_override=}")
num_gpu_blocks = num_gpu_blocks_override
self.cache_config.num_gpu_blocks = num_gpu_blocks
self.cache_config.num_cpu_blocks = num_cpu_blocks
self.model_executor.initialize_cache(num_gpu_blocks, num_cpu_blocks)
@classmethod
def from_engine_args(
cls,