Skip to content

[https://nvbugs/6701493][fix] Latch warmup peak before KV cache size estimation resets it #18669

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
trtllm-agent wants to merge 1 commit into from
Closed

[https://nvbugs/6701493][fix] Latch warmup peak before KV cache size estimation resets it #18669

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
12 changes: 10 additions & 2 deletions tensorrt_llm/_torch/pyexecutor/_util.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1156,6 +1156,14 @@ def configure_kv_cache_capacity(self,

fraction = self._kv_cache_config.free_gpu_memory_fraction

# Warmup (torch.compile specialization, autotuning, CUDA-graph capture,
# memory-pool pre-population) already ran in PyExecutor.__init__ and is
# replayed against the final KV cache once this estimate is applied. Its
# high-water mark can exceed the text-only profiling dummy's, so latch it
# before reset_peak_memory_stats() drops it; otherwise the pool is sized
# against the smaller peak and warmup OOMs on the second pass.
warmup_peak_memory = torch.cuda.max_memory_allocated()

torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
end, total_gpu_memory = torch.cuda.mem_get_info()
Expand Down Expand Up @@ -1197,8 +1205,8 @@ def configure_kv_cache_capacity(self,
if response.has_error():
raise RuntimeError(response.error_msg)

torch_peak_memory = torch.cuda.memory_stats(
)["allocated_bytes.all.peak"]
torch_peak_memory = max(torch.cuda.max_memory_allocated(),
warmup_peak_memory)

# Release before measuring current usage so the retained
# embeddings count toward the peak but not the steady state.
Expand Down
Loading