feat: add three new open clip roberta base models (#860)

* feat: bump openclip to v2.5.0 (#859)

* feat: bump openclip to v2.5.0

* fix: conflicts

* fix: default fp32 on cpu and fp16 on gpu

* feat: add two new models

* fix: remove debug

* fix: add roberta models (test)

* fix: model name xlm

* fix: (wip)

* fix: remove roberta model

* fix: model name

* fix: add :: to model name

* fix: add transformers

* fix: remove is_init_value

* fix: remove transformers

* fix: not use flash-attn on cpu

* fix: add assert description

* fix: housekeeping

* fix: cleanup

* fix: cleanup

* fix: allow to set precision

* fix: gpu-test

* fix: add roberta model test

* fix: dtype

* fix: dtype

* fix: remove optional

* fix: housekeeping

* fix: housekeeping

* fix: typo

* fix: refactor

* fix: refactor

* fix: refactor

* fix: type hint

* fix: address comments

* fix: visiontransformer

* fix: d_model and n_head

* fix: texttransformer

* fix: change model to fix oom gh action

* fix: class init and param name

* fix: black

* chore: bump open-clip version

Co-authored-by: ZiniuYu <[email protected]>

.github/workflows/ci.yml

@@ -113,6 +113,7 @@ jobs:
           pip install --no-cache-dir "server/[onnx]"
           pip install --no-cache-dir "server/[transformers]"
           pip install --no-cache-dir "server/[search]"
+          pip install --no-cache-dir "server/[transformers]"
       - name: Test
         id: test
         run: |

server/clip_server/model/flash_attention.py

@@ -57,9 +57,6 @@ def attention(
             key_padding_mask: a bool tensor of shape (B, S)
 
         """
-        assert not need_weights
-        assert q.dtype in [torch.float16, torch.bfloat16]
-        assert q.is_cuda
 
         if cu_seqlens is None:
             max_s = seqlen