[Bugfix][ROCm] Use batch DMA for CPU KV cache loads (#49843)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Andreas Karatzas
2026-07-27 02:18:01 -07:00
committed by GitHub
co-authored by OpenAI Codex
parent 7f599d7854
commit 394beb633b
2 changed files with 16 additions and 2 deletions
+2 -2
View File
@@ -41,10 +41,10 @@ def _select_swap_blocks_fn(
if gpu_to_cpu:
return ops.swap_blocks_batch
# Fall back to the C++ DMA path on platforms where Triton isn't usable
# (e.g. ROCm builds without Triton) or where GPU kernels cannot directly
# (e.g. ROCm host mappings) or where GPU kernels cannot directly
# dereference CPU pointers (XPU lacks CUDA's unified virtual address space,
# so the Triton kernel's tl.load(cpu_ptr) is invalid on XPU).
if not HAS_TRITON or current_platform.is_xpu():
if not HAS_TRITON or current_platform.is_xpu() or current_platform.is_rocm():
return ops.swap_blocks_batch
page_sizes = [r.page_size_bytes for g in kv_cache_groups_data_refs for r in g]
# Triton wins only on small, 8-byte-aligned payloads.