🧹 Cleanup of the batch size environment variables (#121)

* refactor HF_BATCH_SIZE and BATCH_SIZE into MAX_BATCH_SIZE

* change default batch size to 4

Author: baptistecolle

optimum/tpu/model.py

@@ -10,7 +10,7 @@
 
 
 def get_export_kwargs_from_env():
-    batch_size = os.environ.get("HF_BATCH_SIZE", None)
+    batch_size = os.environ.get("MAX_BATCH_SIZE", None)
     if batch_size is not None:
         batch_size = int(batch_size)
     sequence_length = os.environ.get("HF_SEQUENCE_LENGTH", None)

text-generation-inference/README.md

@@ -65,7 +65,7 @@ docker run -p 8080:80 \
        --net=host --privileged \
        -v $(pwd)/data:/data \
        -e HF_TOKEN=${HF_TOKEN} \
-       -e HF_BATCH_SIZE=1 \
+       -e MAX_BATCH_SIZE=4 \
        -e HF_SEQUENCE_LENGTH=1024 \
        ghcr.io/huggingface/tpu-tgi:latest \
        --model-id mistralai/Mistral-7B-v0.1 \

text-generation-inference/docker/entrypoint.sh

@@ -5,10 +5,10 @@
 ulimit -l 68719476736
 
 # Hugging Face Hub related
-if [[ -z "${BATCH_SIZE}" ]]; then
-  BATCH_SIZE=2
+if [[ -z "${MAX_BATCH_SIZE}" ]]; then
+  MAX_BATCH_SIZE=4
 fi
-export BATCH_SIZE="${BATCH_SIZE}"
+export MAX_BATCH_SIZE="${MAX_BATCH_SIZE}"
 
 if [[ -z "${JSON_OUTPUT_DISABLE}" ]]; then
   JSON_OUTPUT_DISABLE=--json-output
@@ -33,6 +33,6 @@ export QUANTIZATION="${QUANTIZATION}"
 
 
 exec text-generation-launcher --port 8080 \
-  --max-batch-size ${BATCH_SIZE} \
+  --max-batch-size ${MAX_BATCH_SIZE} \
   ${JSON_OUTPUT_DISABLE} \
   --model-id ${MODEL_ID}

text-generation-inference/integration-tests/conftest.py

@@ -109,7 +109,7 @@ def docker_launcher(
         if HUGGING_FACE_HUB_TOKEN is not None:
             env["HUGGING_FACE_HUB_TOKEN"] = HUGGING_FACE_HUB_TOKEN
 
-        for var in ["HF_BATCH_SIZE", "HF_SEQUENCE_LENGTH"]:
+        for var in ["MAX_BATCH_SIZE", "HF_SEQUENCE_LENGTH"]:
             if var in os.environ:
                 env[var] = os.environ[var]
 

text-generation-inference/server/text_generation_server/cli.py

@@ -75,7 +75,7 @@ def serve(
     from .server import serve
 
     # Read environment variables forwarded by the launcher
-    max_batch_size = int(os.environ.get("MAX_BATCH_SIZE", "1"))
+    max_batch_size = int(os.environ.get("MAX_BATCH_SIZE", "4"))
     max_total_tokens = int(os.environ.get("MAX_TOTAL_TOKENS", "64"))
 
     # Start the server