vllm-project/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-12-06 15:04:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

[ROCm][CI] Increase the memory threshold for test_deep_sleep_fp8_kvcache (#30104)

Browse Source 
Signed-off-by: charlifu <charlifu@amd.com>
This commit is contained in:
Charlie Fu
2025-12-04 22:51:44 -06:00
committed by GitHub
parent 5867819eaf
commit 2c22c4ca2d
1 changed files with 7 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
tests/basic_correctness/test_cumem.py
View File

@@ -260,13 +260,18 @@ def test_deep_sleep_fp8_kvcache():
llm.sleep(level=2)
used_bytes = current_platform.get_current_memory_usage() - used_bytes_baseline
assert used_bytes < 3 * GiB_bytes
# Rocm uses more memory for CudaGraphs, so we add 2 GiB more for the threshold
rocm_extra_mem_bytes = 2 * GiB_bytes if current_platform.is_rocm() else 0
mem_threshold_after_sleep = 3 * GiB_bytes + rocm_extra_mem_bytes
assert used_bytes < mem_threshold_after_sleep
llm.wake_up(tags=["weights"])
llm.collective_rpc("reload_weights")
used_bytes = current_platform.get_current_memory_usage() - used_bytes_baseline
assert used_bytes < 4 * GiB_bytes
mem_threshold_after_wake_up = 4 * GiB_bytes + rocm_extra_mem_bytes
assert used_bytes < mem_threshold_after_wake_up
# now allocate kv cache and cuda graph memory
llm.wake_up(tags=["kv_cache"])