Add huggingface/pytorch/tei/docker/1.9.3

[alvarobartt]

Description of changes:

This PR adds the Text Embeddings Inference (TEI) v1.9.3 container to be released on both AWS SageMaker and AWS EC2, as per the latest release of the upstream in https://github.com/huggingface/text-embeddings-inference/releases/v1.9.3.

This PR is pretty similar to earlier releases as #141 or #153, with some subtle differences:

• The entrypoint.sh no longer lives on https://github.com/huggingface/text-embeddings-inference, but rather in this repository instead
• The text-embeddings-router command in the entrypoint.sh no longer requires the --port 8080 flag, as the environment variable PORT has been updated to be set to 8080 instead of 80.
• The entrypoint.sh for both CPU and NVIDIA GPUs runs the text-embeddings-router binary with exec so that the signals are captured by the process, as otherwise the running PID of text-embeddings-router would be different to 1, hence wouldn't capture the signals
• Both Dockerfile files drop the gRPC layer as it wasn't used but it was being compiled / built nonetheless

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Note

By submitting this PR, I disclose that all the code in this PR was written entirely by me, @alvarobartt, without the use of any coding assistants or third-party agentic tools.

[Jyothirmaikottu]

The final stage is named AS base but the build system hardcodes --target sagemaker (release_utils.py:390). This is what causes the CI failure. Previous versions use FROM base AS sagemaker as the final stage — please add that here too, or rename AS base to AS sagemaker.

Affects both 1.9.3/gpu/Dockerfile (line 109) and 1.9.3/cpu/Dockerfile (line 64).

[Jyothirmaikottu]

TEI 1.9.3 SM Test Failure — Root Cause Analysis

The CannotStartContainerError on the BAAI/bge-m3 test is caused by two issues in the new entrypoint.sh:

Issue 1: SageMaker passes serve as CMD argument

SageMaker launches containers with docker run <image> serve. In the 1.9.3 entrypoint, this gets passed through $@ to the router binary:

exec text-embeddings-router-80 "$@"   # $@ = "serve" from SageMaker

text-embeddings-router doesn't understand serve as an argument → immediate crash.

Issue 2: Missing --port 8080

Even without the serve issue, the router doesn't get --port 8080. The PORT=8080 env var in the Dockerfile is not read by the router binary — it needs the CLI flag explicitly. SageMaker health-checks on port 8080, gets no response, marks the container as failed.

Compare with working 1.8.2

The 1.8.2 entrypoint (sagemaker-entrypoint-cuda-all.sh) hardcodes the args:

text-embeddings-router-80 --port 8080 --json-output

It never uses $@, so SageMaker's serve argument is safely ignored.

Fix

Update huggingface/pytorch/tei/docker/1.9.3/gpu/entrypoint.sh — replace each exec line:

# Before (broken):
exec text-embeddings-router-80 "$@"

# After (fixed):
exec text-embeddings-router-80 --port 8080 --json-output

Apply the same to all compute cap branches (75, 80, 90, 100, 120) and the CPU entrypoint.

The CMD ["--json-output"] in the Dockerfile can then be removed since the entrypoint hardcodes the args.