Skip to content

kvcache: skip multi-turn cache reads in decode-only mode #284

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
LouisDDN wants to merge 1 commit into
base: main
Choose a base branch
from
Open

kvcache: skip multi-turn cache reads in decode-only mode #284

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 11 additions & 11 deletions kv_cache_benchmark/kv_cache/benchmark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -514,19 +514,19 @@ def process_requests(self, stop_event: threading.Event):
storage_latency += read_lat
request.context_tokens = remaining_tokens

# 2. For multi-turn conversations, access cache from previous turn.
if self.conversation_manager and request.turn_number > 1:
prev_turn_key = f"{request.conversation_id}_turn_{request.turn_number - 1}"
location, read_latency = self.cache.access_cache(prev_turn_key, InferencePhase.DECODE, 'multi_turn')
if location is not None:
storage_latency += read_latency
with self.results_lock: self.results['multi_turn_cache_hits'] += 1
else:
with self.results_lock: self.results['multi_turn_cache_misses'] += 1

# 3. Perform the main PREFILL operation (a cache WRITE).
# Skip if decode_only mode (disaggregated decode node)
if not self.decode_only:
# 2. For multi-turn conversations, access cache from previous turn.
if self.conversation_manager and request.turn_number > 1:
prev_turn_key = f"{request.conversation_id}_turn_{request.turn_number - 1}"
location, read_latency = self.cache.access_cache(prev_turn_key, InferencePhase.DECODE, 'multi_turn')
if location is not None:
storage_latency += read_latency
with self.results_lock: self.results['multi_turn_cache_hits'] += 1
else:
with self.results_lock: self.results['multi_turn_cache_misses'] += 1

# 3. Perform the main PREFILL operation (a cache WRITE).
if request.phase == InferencePhase.PREFILL or request.phase == InferencePhase.PREFILL_DECODE:
success, location, write_latency = self.cache.allocate_cache(
request.cache_key, request.context_tokens, InferencePhase.PREFILL
Expand Down
Loading