Skip to content

[https://nvbugs/6330273][fix] In StorageManager.__init__, when typical_batch is supplied, append a synthetic… #15465

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

Closed
tensorrt-cicd wants to merge 1 commit into from
+10 −1
Closed

[https://nvbugs/6330273][fix] In StorageManager.__init__, when typical_batch is supplied, append a synthetic… #15465

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
11 changes: 10 additions & 1 deletion tensorrt_llm/runtime/kv_cache_manager_v2/_storage_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,8 +222,17 @@ def __init__(
gpu_quota = config.cache_tiers[GPU_LEVEL].quota
gpu_granularity = CacheLevelManager.cache_tier_granularity(CacheTier.GPU_MEM, gpu_quota)

# Synthesize a constraint from typical_batch so windowed pool groups
# (window_size < tokens_per_block) don't collapse to min_slots = 1 and
# deadlock the scheduler at concurrency > 1. KVCacheDesc(tokens_per_block,
# tokens_per_block - 1) yields non_stale = 1 per request in every pool
# group, floor-ing min_slots at the concurrency level.
effective_constraints = list(constraints or [])
if typical_batch is not None and typical_batch.kv_caches:
one_decode = KVCacheDesc(capacity=tokens_per_block, history_length=tokens_per_block - 1)
effective_constraints.append(BatchDesc([one_decode] * len(typical_batch.kv_caches)))
self._min_slots = self._compute_min_slots_from_constraints(
constraints or [], tokens_per_block, swa_scratch_reuse
effective_constraints, tokens_per_block, swa_scratch_reuse
)

# Compute init_ratio from typical_batch, constraints, or fallback.
Expand Down
Loading