Skip to content

[https://nvbugs/6701493][fix] Limit Gemma3 FP8 accuracy test KV cache memory fraction #18523

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
erictsai-nv wants to merge 4 commits into from
Closed

[https://nvbugs/6701493][fix] Limit Gemma3 FP8 accuracy test KV cache memory fraction #18523

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
b56cbbe
[https://nvbugs/6701493][fix] Limit Gemma3 FP8 accuracy test KV cache…
erictsai-nv Sep 1, 2026
53dcb7e
Update tests/integration/defs/accuracy/test_llm_api_pytorch.py
erictsai-nv Sep 2, 2026
87f5d10
Fix duplicate dtype assignment in kv_cache_config
erictsai-nv Sep 2, 2026
2b47250
Fix Gemma3 FP8 test formatting
erictsai-nv Sep 2, 2026
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
5 changes: 4 additions & 1 deletion tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1400,9 +1400,12 @@ def test_auto_dtype(self):

def test_fp8_prequantized(self):
# Disabling kv cache reuse as a WAR to deal with gaps in kernel support for Gemma3's non-inclusive sliding window size.
# Cap KV cache at 50% of free memory to leave PyTorch
# activation headroom during MMLU (nvbugs/6701493).
kv_cache_config = KvCacheConfig(enable_block_reuse=False,
enable_partial_reuse=False,
dtype="fp8")
dtype="fp8",
free_gpu_memory_fraction=0.5)
Comment thread
erictsai-nv marked this conversation as resolved.
# Note: This has only the LLM part quantized. Vision part is in bfloat16.
prequantized_model_path = f"{llm_models_root()}/gemma/gemma-3-27b-it-fp8/"
with LLM(prequantized_model_path,
Expand Down
Loading