[ROCm]: Bump torch 2.12, triton 3.7, torchaudio, torchvision (#50607)

Signed-off-by: Rohan Potdar <rohan.potdar@amd.com>
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Rohan Potdar
2026-08-04 20:14:11 -05:00
committed by GitHub
co-authored by Claude Opus 4.8
parent 43c4bdcae9
commit d0ce3dadb6
2 changed files with 21 additions and 14 deletions
+11 -3
View File
@@ -22,6 +22,7 @@ from vllm.compilation.counter import compilation_counter
from vllm.config import VllmConfig, get_current_vllm_config
from vllm.config.utils import hash_factors
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils.hashing import safe_hash
try:
@@ -130,9 +131,16 @@ class StandaloneCompiledArtifacts:
compilation_counter.num_compiled_artifacts_loaded += 1
return AOTCompiledArtifact.deserialize(entry)
with concurrent.futures.ThreadPoolExecutor() as executor:
entries = list(self.submodule_bytes_store.values())
loaded_entries = list(executor.map(_load_entry, entries))
entries = list(self.submodule_bytes_store.values())
if current_platform.is_rocm():
# Deserializing an artifact loads its compiled Triton kernels, which
# on ROCm ends up in hipModuleLoad. Loading modules concurrently
# deadlocks on the dynamic linker lock, since rocprofiler-sdk calls
# dl_iterate_phdr from inside the load, so load one at a time.
loaded_entries = [_load_entry(entry) for entry in entries]
else:
with concurrent.futures.ThreadPoolExecutor() as executor:
loaded_entries = list(executor.map(_load_entry, entries))
for i, k in enumerate(self.submodule_bytes_store.keys()):
self.loaded_submodule_store[k] = loaded_entries[i]