Run:ai model streamer add GCS package support (#24909)

Signed-off-by: Peter Schuurman <psch@google.com>

docs/models/extensions/runai_model_streamer.md

@@ -24,6 +24,13 @@ vllm serve s3://core-llm/Llama-3-8b \
     --load-format runai_streamer
 ```
 
+To run model from Google Cloud Storage run:
+
+```bash
+vllm serve gs://core-llm/Llama-3-8b \
+    --load-format runai_streamer
+```
+
 To run model from a S3 compatible object store run:
 
 ```bash

requirements/nightly_torch_test.txt

@@ -43,6 +43,6 @@ tritonclient==2.51.0
 numba == 0.60.0; python_version == '3.9' # v0.61 doesn't support Python 3.9. Required for N-gram speculative decoding
 numba == 0.61.2; python_version > '3.9'
 numpy
-runai-model-streamer[s3]==0.14.0
+runai-model-streamer[s3,gcs]==0.14.0
 fastsafetensors>=0.1.10
 pydantic>=2.10 # 2.9 leads to error on python 3.10

requirements/rocm.txt

@@ -13,6 +13,6 @@ tensorizer==2.10.1
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 setuptools-scm>=8
-runai-model-streamer[s3]==0.14.0
+runai-model-streamer[s3,gcs]==0.14.0
 conch-triton-kernels==1.2.1
-timm>=1.0.17
\ No newline at end of file
+timm>=1.0.17

requirements/test.in

@@ -51,7 +51,7 @@ tritonclient==2.51.0
 numba == 0.60.0; python_version == '3.9' # v0.61 doesn't support Python 3.9. Required for N-gram speculative decoding
 numba == 0.61.2; python_version > '3.9'
 numpy
-runai-model-streamer[s3]==0.14.0
+runai-model-streamer[s3,gcs]==0.14.0
 fastsafetensors>=0.1.10
 pydantic>=2.10 # 2.9 leads to error on python 3.10
 decord==0.6.0

requirements/test.txt

@@ -251,11 +251,27 @@ gitdb==4.0.12
 gitpython==3.1.44
     # via mlflow-skinny
 google-api-core==2.24.2
-    # via opencensus
+    # via
+    #   google-cloud-core
+    #   google-cloud-storage
+    #   opencensus
 google-auth==2.40.2
     # via
     #   databricks-sdk
     #   google-api-core
+    #   google-cloud-core
+    #   google-cloud-storage
+    #   runai-model-streamer-gcs
+google-cloud-core==2.4.3
+    # via google-cloud-storage
+google-cloud-storage==3.4.0
+    # via runai-model-streamer-gcs
+google-crc32c==1.7.1
+    # via
+    #   google-cloud-storage
+    #   google-resumable-media
+google-resumable-media==2.7.2
+    # via google-cloud-storage
 googleapis-common-protos==1.70.0
     # via google-api-core
 graphene==3.4.3
@@ -890,6 +906,7 @@ requests==2.32.3
     #   docker
     #   evaluate
     #   google-api-core
+    #   google-cloud-storage
     #   huggingface-hub
     #   lightly
     #   lm-eval
@@ -929,6 +946,8 @@ rtree==1.4.0
     # via torchgeo
 runai-model-streamer==0.14.0
     # via -r requirements/test.in
+runai-model-streamer-gcs==0.14.0
+    # via runai-model-streamer
 runai-model-streamer-s3==0.14.0
     # via runai-model-streamer
 s3transfer==0.10.3

tests/model_executor/model_loader/runai_model_streamer/test_runai_utils.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import glob
+import hashlib
 import os
 import tempfile
 
@@ -9,7 +10,8 @@ import huggingface_hub.constants
 
 from vllm.model_executor.model_loader.weight_utils import (
     download_weights_from_hf)
-from vllm.transformers_utils.runai_utils import (is_runai_obj_uri,
+from vllm.transformers_utils.runai_utils import (ObjectStorageModel,
+                                                 is_runai_obj_uri,
                                                  list_safetensors)
 
 
@@ -34,6 +36,23 @@ def test_runai_list_safetensors_local():
         assert len(safetensors) == len(files)
 
 
-if __name__ == "__main__":
-    test_is_runai_obj_uri()
-    test_runai_list_safetensors_local()
+def test_runai_pull_files_gcs(monkeypatch):
+    monkeypatch.setenv("RUNAI_STREAMER_GCS_USE_ANONYMOUS_CREDENTIALS", "true")
+    # Bypass default project lookup by setting GOOGLE_CLOUD_PROJECT
+    monkeypatch.setenv("GOOGLE_CLOUD_PROJECT", "fake-project")
+    filename = "LT08_L1GT_074061_20130309_20170505_01_T2_MTL.txt"
+    gcs_bucket = "gs://gcp-public-data-landsat/LT08/01/074/061/LT08_L1GT_074061_20130309_20170505_01_T2/"
+    gcs_url = f"{gcs_bucket}/{filename}"
+    model = ObjectStorageModel(gcs_url)
+    model.pull_files(gcs_bucket, allow_pattern=[f"*{filename}"])
+    # To re-generate / change URLs:
+    #   gsutil ls -L gs://<gcs-url> | grep "Hash (md5)" | tr -d ' ' \
+    #     | cut -d":" -f2 | base64 -d | xxd -p
+    expected_checksum = "f60dea775da1392434275b311b31a431"
+    hasher = hashlib.new("md5")
+    with open(os.path.join(model.dir, filename), 'rb') as f:
+        # Read the file in chunks to handle large files efficiently
+        for chunk in iter(lambda: f.read(4096), b''):
+            hasher.update(chunk)
+    actual_checksum = hasher.hexdigest()
+    assert actual_checksum == expected_checksum