[V1] Support any head size for FlexAttention backend (#20467)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/models/multimodal/generation/test_common.py

@@ -33,9 +33,6 @@ if current_platform.is_rocm():
     os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
 
 REQUIRES_V0_MODELS = [
-    # V1 Test: no way to fall back for head_dim = 80
-    # https://github.com/vllm-project/vllm/issues/14524
-    "qwen_vl",
     # V1 Test: not enough KV cache space in C1.
     "fuyu",
 ]
@@ -221,8 +218,7 @@ VLM_TEST_SETTINGS = {
         marks=[large_gpu_mark(min_gb=32)],
     ),
     "blip2": VLMTestInfo(
-        # TODO: Change back to 2.7b once head_dim = 80 is supported
-        models=["Salesforce/blip2-opt-6.7b"],
+        models=["Salesforce/blip2-opt-2.7b"],
         test_type=VLMTestType.IMAGE,
         prompt_formatter=lambda img_prompt: f"Question: {img_prompt} Answer:",
         img_idx_to_prompt=lambda idx: "",
@@ -340,8 +336,7 @@ VLM_TEST_SETTINGS = {
     "h2ovl": VLMTestInfo(
         models = [
             "h2oai/h2ovl-mississippi-800m",
-            # TODO: Re-enable once head_dim = 80 is supported
-            # "h2oai/h2ovl-mississippi-2b",
+            "h2oai/h2ovl-mississippi-2b",
         ],
         test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
         prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>", # noqa: E501

tests/models/quantization/test_gguf.py

@@ -83,7 +83,7 @@ MODELS = [
     QWEN2_CONFIG,
     PHI3_CONFIG,
     GPT2_CONFIG,
-    # STABLELM_CONFIG,  # enable this when v1 support head_size=80
+    STABLELM_CONFIG,
     DOLPHIN_CONFIG,
     # STARCODER_CONFIG, # broken
 ]

tests/models/registry.py

@@ -240,8 +240,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "OrionForCausalLM": _HfExamplesInfo("OrionStarAI/Orion-14B-Chat",
                                         trust_remote_code=True),
     "PersimmonForCausalLM": _HfExamplesInfo("adept/persimmon-8b-chat"),
-    "PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2", v0_only=True),
+    "PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2"),
     "Phi3ForCausalLM": _HfExamplesInfo("microsoft/Phi-3-mini-4k-instruct"),
+    # Blocksparse attention not supported in V1 yet
     "Phi3SmallForCausalLM": _HfExamplesInfo("microsoft/Phi-3-small-8k-instruct",
                                             trust_remote_code=True,
                                             v0_only=True),
@@ -258,10 +259,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
     "Qwen3ForSequenceClassification": _HfExamplesInfo("tomaarsen/Qwen3-Reranker-0.6B-seq-cls"),  # noqa: E501
     "RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
-    "StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b",  # noqa: E501
-                                                v0_only=True),
-    "StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t",
-                                           v0_only=True),
+    "StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b"),  # noqa: E501
+    "StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
     "Starcoder2ForCausalLM": _HfExamplesInfo("bigcode/starcoder2-3b"),
     "SolarForCausalLM": _HfExamplesInfo("upstage/solar-pro-preview-instruct"),
     "TeleChat2ForCausalLM": _HfExamplesInfo("Tele-AI/TeleChat2-3B",
@@ -330,8 +329,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
     "AyaVisionForConditionalGeneration": _HfExamplesInfo("CohereForAI/aya-vision-8b"), # noqa: E501
     "Blip2ForConditionalGeneration": _HfExamplesInfo("Salesforce/blip2-opt-2.7b",  # noqa: E501
-                                                     extras={"6b": "Salesforce/blip2-opt-6.7b"},  # noqa: E501
-                                                     v0_only=True),
+                                                     extras={"6b": "Salesforce/blip2-opt-6.7b"}),  # noqa: E501
     "ChameleonForConditionalGeneration": _HfExamplesInfo("facebook/chameleon-7b"),  # noqa: E501
     "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny",  # noqa: E501
                                                 extras={"fork": "Isotr0py/deepseek-vl2-tiny"},  # noqa: E501
@@ -359,8 +357,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
                                                     trust_remote_code=True),
     "KimiVLForConditionalGeneration": _HfExamplesInfo("moonshotai/Kimi-VL-A3B-Instruct",  # noqa: E501
                                                       extras={"thinking": "moonshotai/Kimi-VL-A3B-Thinking"},  # noqa: E501
-                                                      trust_remote_code=True,
-                                                      v0_only=True),
+                                                      trust_remote_code=True),
     "Llama4ForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct",   # noqa: E501
                                                       max_model_len=10240),
     "LlavaForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-1.5-7b-hf",

tests/models/test_initialization.py

@@ -22,7 +22,8 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
     model_info.check_transformers_version(on_fail="skip")
 
     # FIXME: Possible memory leak in the previous tests?
-    if model_arch == "GraniteSpeechForConditionalGeneration":
+    if model_arch in ("GraniteSpeechForConditionalGeneration",
+                      "KimiVLForConditionalGeneration"):
         pytest.skip("Avoid OOM")
 
     # Avoid OOM and reduce initialization time by only using 1 layer