A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 19:30 UTC
[Hardware][AMD][CI][Bugfix] Fix regressions from deprecated env vars (#32837)
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
Matt committed
Jan 22, 2026 at 10:59 UTC
c517d8c93404738effe29b5beea606111aba9258
3 files changed
+14
-6
.buildkite/test-amd.yaml
+2
-2
@@ -1475,7 +1475,7 @@ steps:
1475
- tests/v1/kv_connector/nixl_integration/
1476
commands:
1477
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1478
- - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh --attention-backend ROCM_ATTN
1478
+ - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1479
1480
- label: DP EP NixlConnector PD accuracy tests (Distributed) # 15min
1481
mirror_hardwares: [amdexperimental, amdproduction]
@@ -1489,7 +1489,7 @@ steps:
1489
- tests/v1/kv_connector/nixl_integration/
1490
commands:
1491
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1492
- - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh --attention-backend ROCM_ATTN
1492
+ - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1493
1494
##### multi gpus test #####
1495
##### A100 test #####
tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+6
-1
@@ -43,7 +43,12 @@ run_tests() {
43
}
44
45
# Run tests
46
-run_tests "default backend" ""
46
+if [[ -n "${ROCM_ATTN:-}" ]]; then
47
+ echo "ROCM_ATTN is set, running with --attention-backend ROCM_ATTN"
48
+ run_tests "ROCM_ATTN backend" "--attention-backend ROCM_ATTN"
49
+else
50
+ run_tests "default backend" ""
51
+fi
52
53
# Check if FLASHINFER is set (non-empty)
54
if [[ -n "${FLASHINFER:-}" ]]; then
vllm/platforms/rocm.py
+6
-3
@@ -293,10 +293,13 @@ class RocmPlatform(Platform):
293
return AttentionBackendEnum.ROCM_AITER_FA.get_path()
294
295
# Priority 3: Check for ROCM_ATTN (prefill-decode split)
296
- from vllm.config import get_current_vllm_config
296
+ from vllm.config import get_current_vllm_config_or_none
297
298
- vllm_config = get_current_vllm_config()
299
- if vllm_config.attention_config.use_prefill_decode_attention:
298
+ vllm_config = get_current_vllm_config_or_none()
299
+ if (
300
+ vllm_config is not None
301
+ and vllm_config.attention_config.use_prefill_decode_attention
302
+ ):
303
logger.info("Using Rocm Attention backend.")
304
return AttentionBackendEnum.ROCM_ATTN.get_path()
305