[Frontend][4/N] Improve all pooling task | Add plugin pooling task (#26973)

Signed-off-by: wang.yuqi <noooop@126.com>
Signed-off-by: Christian Pinto <christian.pinto@ibm.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Christian Pinto <christian.pinto@ibm.com>
This commit is contained in:
wang.yuqi
2025-10-23 14:46:18 +00:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Christian Pinto
parent fe2016de2d
commit 3fa2c12185
16 changed files with 102 additions and 54 deletions
@@ -40,7 +40,7 @@ def _run_test(
max_num_seqs=32,
default_torch_num_threads=1,
) as vllm_model:
vllm_model.llm.encode(prompt, pooling_task="token_classify")
vllm_model.llm.encode(prompt, pooling_task="plugin")
MODELS = ["mgazz/Prithvi-EO-2.0-300M-TL-Sen1Floods11"]
@@ -368,9 +368,9 @@ class PrithviMultimodalDataProcessor(IOProcessor):
out_format = "b64_json"
for output in model_output:
y_hat = output.outputs.data.argmax(dim=1)
y_hat = output.outputs.data.argmax(dim=0)
pred = torch.nn.functional.interpolate(
y_hat.unsqueeze(1).float(),
y_hat[None, None, ...].float(),
size=self.img_size,
mode="nearest",
)
@@ -9,7 +9,6 @@ from tests.utils import RemoteOpenAIServer
from vllm.config import VllmConfig
from vllm.entrypoints.openai.protocol import IOProcessorResponse
from vllm.plugins.io_processors import get_io_processor
from vllm.pooling_params import PoolingParams
MODEL_NAME = "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11"
@@ -94,8 +93,6 @@ def test_prithvi_mae_plugin_offline(vllm_runner, model_name: str):
out_data_format="b64_json",
)
pooling_params = PoolingParams(activation=False)
with vllm_runner(
model_name,
runner="pooling",
@@ -109,9 +106,7 @@ def test_prithvi_mae_plugin_offline(vllm_runner, model_name: str):
model_impl="terratorch",
io_processor_plugin="prithvi_to_tiff",
) as llm_runner:
pooler_output = llm_runner.get_llm().encode(
img_prompt, pooling_params=pooling_params, pooling_task="token_classify"
)
pooler_output = llm_runner.get_llm().encode(img_prompt, pooling_task="plugin")
output = pooler_output[0].outputs
# verify the output is formatted as expected for this plugin