fix(server): Fix Transformers fork version

.dockerignore

 aml
 target
+server/transformers
\ No newline at end of file

Dockerfile

@@ -2,6 +2,7 @@ FROM rust:1.64 as router-builder
 
 WORKDIR /usr/src
 
+COPY rust-toolchain.toml rust-toolchain.toml
 COPY proto proto
 COPY router router
 
@@ -13,6 +14,7 @@ FROM rust:1.64 as launcher-builder
 
 WORKDIR /usr/src
 
+COPY rust-toolchain.toml rust-toolchain.toml
 COPY launcher launcher
 
 WORKDIR /usr/src/launcher

aml/deployment.yaml

@@ -8,7 +8,7 @@ environment_variables:
   MODEL_NAME: bigscience/bloom
   NUM_GPUS: 8
 environment:
-  image: db4c2190dd824d1f950f5d1555fbadf0.azurecr.io/text-generation-inference:0.2
+  image: db4c2190dd824d1f950f5d1555fbadf0.azurecr.io/text-generation-inference:0.3.1
   inference_config:
     liveness_route:
       port: 3000
@@ -25,14 +25,14 @@ request_settings:
   max_concurrent_requests_per_instance: 256
 liveness_probe:
   initial_delay: 600
-  timeout: 20
+  timeout: 90
   period: 120
   success_threshold: 1
-  failure_threshold: 3
+  failure_threshold: 5
 readiness_probe:
   initial_delay: 600
-  timeout: 20
+  timeout: 90
   period: 120
   success_threshold: 1
-  failure_threshold: 3
+  failure_threshold: 5
 instance_count: 1

server/Makefile

@@ -7,13 +7,13 @@ gen-server:
 	touch text_generation/pb/__init__.py
 
 install-transformers:
-	# Install specific version of transformers
+	# Install specific version of transformers with custom cuda kernels
 	rm transformers || true
-	rm transformers-7302a24535e8dc5637ea5b4e4572fc971d404098 || true
-	curl -L -O https://github.com/OlivierDehaene/transformers/archive/7302a24535e8dc5637ea5b4e4572fc971d404098.zip
-	unzip 7302a24535e8dc5637ea5b4e4572fc971d404098.zip
-	rm 7302a24535e8dc5637ea5b4e4572fc971d404098.zip
-	mv transformers-7302a24535e8dc5637ea5b4e4572fc971d404098 transformers
+	rm transformers-b55f16c5b71aeef47a66a4270e19c154f050a7a7 || true
+	curl -L -O https://github.com/OlivierDehaene/transformers/archive/b55f16c5b71aeef47a66a4270e19c154f050a7a7.zip
+	unzip b55f16c5b71aeef47a66a4270e19c154f050a7a7.zip
+	rm b55f16c5b71aeef47a66a4270e19c154f050a7a7.zip
+	mv transformers-b55f16c5b71aeef47a66a4270e19c154f050a7a7 transformers
 	cd transformers && python setup.py install
 
 install-torch:

server/text_generation/models/bloom.py

@@ -38,7 +38,7 @@ class BLOOMSharded(CausalLM):
         self.master = self.rank == 0
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{self.rank}")
-            dtype = torch.float16
+            dtype = torch.bfloat16
         else:
             device = torch.device("cpu")
             dtype = torch.float32