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 11, 00:00 UTC
[ROCm][CI][Bugfix] Fix Siglip2 rotary embedding dispatch and InternVL video test tolerance (#31235)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Dec 23, 2025 at 20:56 UTC
e42894f5b530c12b4783ec37d6b5b345b9ccb349
2 files changed
+4
-1
tests/models/multimodal/generation/test_common.py
+1
@@ -513,6 +513,7 @@ VLM_TEST_SETTINGS = {
513
max_model_len=8192,
514
use_tokenizer_eos=True,
515
patch_hf_runner=model_utils.internvl_patch_hf_runner,
516
+ num_logprobs=10 if current_platform.is_rocm() else 5,
517
),
518
"intern_vl-hf": VLMTestInfo(
519
models=["OpenGVLab/InternVL3-1B-hf"],
vllm/model_executor/models/siglip2navit.py
+3
-1
@@ -163,8 +163,10 @@ def apply_rotary_pos_emb(
163
enable_fp32_compute=True,
164
)
165
166
- if is_flash_attn_backend and not current_platform.is_cuda():
166
+ if is_flash_attn_backend and current_platform.is_cuda():
167
apply_rotary_emb_func = apply_rotary_emb.forward_cuda
168
+ elif is_flash_attn_backend and current_platform.is_rocm():
169
+ apply_rotary_emb_func = apply_rotary_emb.forward_hip
170
else:
171
apply_rotary_emb_func = apply_rotary_emb.forward_native
172