EmbeddingGemma 2

✓ Verified by Badgr · Transformers · Run

Last verified 2026-10-10

badgr run --image vllm/vllm-openai:v0.30.0 --gpu RTX_4090 --funding paid --max-cost 1 --max-runtime 25 --cmd 'pip install sentence-transformers ... python: SentenceTransformer("google/embeddinggemma-2").encode(...)'

GPU

NVIDIA RTX A5000Reported by the provider, not confirmed inside the container

Command runtime

290s

Deployment

dep-6adff26a5a

• ✓ Workload container started
• ✓ Command exited with code 0
• ✓ Embeddings returned (dimension 768)
• ✓ Teardown completed

Run through sentence-transformers rather than vLLM: a vLLM 0.31 nightly fails to start this model on 99 KB-shared-memory GPUs (vllm-project/vllm#60847). The request asked for an RTX 4090 and the job reported an RTX A5000. The command is abridged and the elapsed time is estimated from cost and hourly rate.

GitHub issue this run was for: vllm-project/vllm#60847

deployment=dep-6adff26a5a status=succeeded user_command_exit=0
gpu NVIDIA RTX A5000 | dim 768 | load+encode 11.2s
sim(query, relevant) = 0.871 | sim(query, unrelated) = 0.577

Run from Badgr's local development environment. Documents that this job works end-to-end for this model.

✓ Verified by Badgr · vLLM · Serve

Last verified 2026-10-10

badgr serve --image vllm/vllm-openai:nightly-8cbd5d03006c33185f402249ff2b448efd594986 --gpu A100 --funding paid --port 8000 --health-path /v1/models --max-cost 1 --startup-timeout 25 --cmd "vllm serve google/embeddinggemma-2 --runner pooling --dtype bfloat16 --max-model-len 4096 --limit-mm-per-prompt '{\"image\":0,\"audio\":0,\"video\":0}' --host 0.0.0.0 --port 8000"

GPU

NVIDIA A100-SXM4-40GBConfirmed inside the container

Time to verified

97s

Deployment

dep-3f93200aae

• ✓ vLLM nightly started EmbeddingGemma 2 with --runner pooling on an A100 40GB
• ✓ Badgr verified the endpoint with a real /v1/embeddings request for google/embeddinggemma-2 (768 dimensions)
• ✓ A two-sentence embeddings request returned unit-norm 768-dimension vectors
• ✓ final_state = VERIFIED
• ✓ Teardown completed

Upstream vllm#60847: this model fails at startup on consumer Ampere/Blackwell cards (99 KiB shared memory per block) and an RTX 3090 reproduced it; the A100 is not affected, so Badgr is routed to one. The first attempt was marked failed by Badgr itself, which checked /v1/completions with the default model name on a server that only has embeddings. Both are fixed: Badgr now verifies a raw vLLM command with the model it serves and an embedding server with /v1/embeddings. Serve-verified with the issue's nightly image.

GitHub issue this run was for: vllm-project/vllm#60847

final_state=VERIFIED failure_class=None providers_tried=1/3 teardown_confirmed=False spend_usd=0.0324

Run from Badgr's local development environment. Documents that this deployment path works end-to-end for this model.