use transformers from git

Dockerfile

@@ -31,7 +31,6 @@ ENV PDM_CACHE_DIR $HOME/code/.pdm_cache/
 ENV HF_HOME $HOME/code/.hf_home/
 
 ENV DEVICE cuda:0
-# ENV ATTN_IMPLEMENTATION flash_attention_2
-ENV ATTN_IMPLEMENTATION flash_attention
+ENV ATTN_IMPLEMENTATION sdpa
 
 ENTRYPOINT [ "./entrypoint.sh" ]

server/entrypoint.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-curl -sS https://bootstrap.pypa.io/get-pip.py | ./.venv/bin/python
-./.venv/bin/python -m pip install flash_attn==1.0.9 --no-build-isolation
+# curl -sS https://bootstrap.pypa.io/get-pip.py | ./.venv/bin/python
+# ./.venv/bin/python -m pip install flash_attn==1.0.9 --no-build-isolation
 
 echo "Installation done"
 

server/pdm.lock

@@ -5,7 +5,7 @@
 groups = ["default"]
 strategy = ["cross_platform"]
 lock_version = "4.4"
-content_hash = "sha256:4c9d99229e31396fab6ad1bc7aaf336366423f7e144888cb4edb1a4d5dd6a7dc"
+content_hash = "sha256:a7e1894df31010cec5c072c8c7c796e2124f68acffe168327b11e80d5ab0f775"
 
 [[package]]
 name = "accelerate"
@@ -859,8 +859,10 @@ files = [
 
 [[package]]
 name = "transformers"
-version = "4.36.2"
+version = "4.37.0.dev0"
 requires_python = ">=3.8.0"
+git = "https://github.com/huggingface/transformers.git"
+revision = "772307be7649e1333a933cfaa229dc0dec2fd331"
 summary = "State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow"
 dependencies = [
     "filelock",
@@ -874,26 +876,6 @@ dependencies = [
     "tokenizers<0.19,>=0.14",
     "tqdm>=4.27",
 ]
-files = [
-    {file = "transformers-4.36.2-py3-none-any.whl", hash = "sha256:462066c4f74ee52516f12890dcc9ec71d1a5e97998db621668455117a54330f6"},
-    {file = "transformers-4.36.2.tar.gz", hash = "sha256:d8068e897e47793281501e547d2bbdfc5b8556409c2cb6c3d9e2ca77d4c0b4ec"},
-]
-
-[[package]]
-name = "transformers"
-version = "4.36.2"
-extras = ["torch"]
-requires_python = ">=3.8.0"
-summary = "State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow"
-dependencies = [
-    "accelerate>=0.21.0",
-    "torch!=1.12.0,>=1.10",
-    "transformers==4.36.2",
-]
-files = [
-    {file = "transformers-4.36.2-py3-none-any.whl", hash = "sha256:462066c4f74ee52516f12890dcc9ec71d1a5e97998db621668455117a54330f6"},
-    {file = "transformers-4.36.2.tar.gz", hash = "sha256:d8068e897e47793281501e547d2bbdfc5b8556409c2cb6c3d9e2ca77d4c0b4ec"},
-]
 
 [[package]]
 name = "triton"

server/pyproject.toml

@@ -9,9 +9,11 @@ dependencies = [
     "fastapi>=0.109.0",
     "uvicorn[standard]>=0.25.0",
     "numpy>=1.26.3",
-    "transformers[torch]>=4.36.2",
+    "transformers @ git+https://github.com/huggingface/transformers.git",
     "ninja>=1.11.1.1",
     "packaging>=23.2",
+    "torch>=2.1.2",
+    "accelerate>=0.26.1",
 ]
 requires-python = ">=3.11"
 readme = "README.md"