[Bugfix][CI] Gemma3 Transformers multimodal encoder profiling and build prompt-embedding fixtures (#44952)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
@@ -14,6 +14,7 @@ import torch
|
||||
from openai import BadRequestError
|
||||
|
||||
from tests.utils import VLLM_PATH, RemoteOpenAIServer
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODEL_NAME = "facebook/opt-125m"
|
||||
CHAT_TEMPLATE = VLLM_PATH / "examples/template_chatml.jinja"
|
||||
@@ -41,7 +42,11 @@ def server_args() -> list[str]:
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def server(server_args):
|
||||
def server(server_args, request):
|
||||
if current_platform.is_rocm():
|
||||
# Materialize HF embeddings before the server reserves ROCm VRAM.
|
||||
request.getfixturevalue("prompt_embeds_b64")
|
||||
request.getfixturevalue("aligned_content_and_embeds_b64")
|
||||
with RemoteOpenAIServer(MODEL_NAME, server_args) as remote_server:
|
||||
yield remote_server
|
||||
|
||||
|
||||
@@ -72,7 +72,35 @@ class MultiModalProcessingInfo(BaseProcessingInfo):
|
||||
image_sizes=([height, width],), **mm_processor_kwargs
|
||||
)
|
||||
image_tokens = mm_tokens["num_image_tokens"][0]
|
||||
return image_tokens
|
||||
return self._get_max_encoder_tokens(processor, mm_tokens) or image_tokens
|
||||
|
||||
@staticmethod
|
||||
def _get_mm_values(mm_tokens: object, key: str) -> object:
|
||||
if isinstance(mm_tokens, Mapping):
|
||||
return mm_tokens.get(key)
|
||||
return getattr(mm_tokens, key, None)
|
||||
|
||||
def _get_max_encoder_tokens(
|
||||
self, processor: object, mm_tokens: object
|
||||
) -> int | None:
|
||||
if "gemma3" not in processor.__class__.__name__.lower():
|
||||
return None
|
||||
|
||||
vision_config = getattr(self.get_hf_config(), "vision_config", None)
|
||||
image_size = getattr(vision_config, "image_size", None)
|
||||
patch_size = getattr(vision_config, "patch_size", None)
|
||||
if not image_size or not patch_size:
|
||||
return None
|
||||
|
||||
# Gemma3 pools each 64x64 SigLIP patch grid down to 256 image tokens.
|
||||
# Profile the vision encoder against the pre-pooling patch-token count.
|
||||
patches_per_image = (image_size // patch_size) ** 2
|
||||
num_image_patches = self._get_mm_values(mm_tokens, "num_image_patches") or [1]
|
||||
if isinstance(num_image_patches, int):
|
||||
max_image_patches = num_image_patches
|
||||
else:
|
||||
max_image_patches = max(num_image_patches)
|
||||
return patches_per_image * int(max_image_patches)
|
||||
|
||||
def get_max_image_size(self):
|
||||
return 10_000, 10_000 # hardcode for arbitrary very large size
|
||||
|
||||
Reference in New Issue
Block a user