iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

fix(rocm): remove workaround causing invalid argument on Qwen3.5 with TP=2 (#40686)

Browse Source 
Co-authored-by: Test User <test@example.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
Zhaodong Bing
2026-05-06 01:38:32 -07:00
committed by GitHub
co-authored by Test User TJian
parent 1c58876618
commit 66d1cc0c77
1 changed files with 0 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/cumem_allocator.cpp
-22
View File
@@ -232,28 +232,6 @@ void unmap_and_release(unsigned long long device, ssize_t size,
}
}
// ROCm workaround: hipMemRelease does not return physical VRAM to the
// free pool while the virtual-address reservation is still held.
// Cycling cuMemAddressFree → cuMemAddressReserve (at the same address)
// forces the driver to actually release the physical pages while keeping
// the same VA available for a later create_and_map.
if (first_error == no_error) {
first_error = cuMemAddressFree(d_mem, size);
if (first_error == no_error) {
CUdeviceptr d_mem_new = 0;
first_error = cuMemAddressReserve(&d_mem_new, size, 0, d_mem, 0);
if (first_error == no_error && d_mem_new != d_mem) {
cuMemAddressFree(d_mem_new, size);
snprintf(error_msg, sizeof(error_msg),
"ROCm: VA re-reserve got %p instead of %p", (void*)d_mem_new,
(void*)d_mem);
error_code = CUresult(1);
std::cerr << error_msg << std::endl;
return;
}
}
}
if (first_error != no_error) {
CUDA_CHECK(first_error);
}