Skip to content

Conversation

suyoggupta
Copy link

@suyoggupta suyoggupta commented Sep 24, 2025

bench.yaml

# Compilation backend
compile_backend: torch-cudagraph
# Runtime engine
runtime: trtllm
# Model loading
skip_loading_weights: false
# Fraction of free memory to use for kv-caches
free_mem_ratio: 0.9
# CUDA Graph optimization
cuda_graph_batch_sizes: [64, 32,16,8,4,2,1]
# Attention backend
attn_backend: flashinfer
# Sequence configuration
max_batch_size: 64
model_kwargs:
  torch_dtype: bfloat16

Comment on lines +589 to +592
if tnsr_device.numel() < tnsr_host.numel():
print("WARNING: tnsr_device.numel() < tnsr_like.numel()")
print(f"{name=}, {tnsr_device.numel()=}, {tnsr_host.numel()=}")
tnsr_device.resize_(tnsr_host.numel())
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

curious where this is necessary?

Comment on lines +589 to +592
if tnsr_device.numel() < tnsr_host.numel():
print("WARNING: tnsr_device.numel() < tnsr_like.numel()")
print(f"{name=}, {tnsr_device.numel()=}, {tnsr_host.numel()=}")
tnsr_device.resize_(tnsr_host.numel())
Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@lucaslie : FYI, this is the WAR I have to get resize functionality working again on the feature branch. Without this llama3.1 + cache_resize is broken

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants