biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Fix DeepSeek V3.2 OOM during CG memory profiling (#36691)

Browse Source 
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2026-03-10 23:01:07 -04:00
committed by GitHub
parent 84e436ed1c
commit 8ab3d7427c
1 changed files with 5 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/v1/worker/gpu_model_runner.py
View File

@@ -5550,16 +5550,14 @@ class GPUModelRunner(
kv_cache_spec = self.get_kv_cache_spec()
kv_cache_groups = get_kv_cache_groups(self.vllm_config, kv_cache_spec)
min_blocks = self.compilation_config.max_cudagraph_capture_size or 1
if kv_cache_groups:
page_size = kv_cache_groups[0].kv_cache_spec.page_size_bytes
group_size = max(len(g.layer_names) for g in kv_cache_groups)
available_memory = min_blocks * page_size * group_size
else:
available_memory = 1 # Attention-free model
# Temporarily change num_gpu_blocks_override to allocate a minimal KV cache
saved_override = self.cache_config.num_gpu_blocks_override
self.cache_config.num_gpu_blocks_override = min_blocks
minimal_config = get_kv_cache_config_from_groups(
self.vllm_config, kv_cache_groups, available_memory=available_memory
self.vllm_config, kv_cache_groups, available_memory=0
)
self.cache_config.num_gpu_blocks_override = saved_override
self.initialize_kv_cache(minimal_config)
self.cache_config.num_gpu_blocks = minimal_config.num_blocks