On the Bedrock OpenAI-compatible chat completions endpoint, repeated identical requests to google.gemma-4-26b-a4b at temperature 0 (greedy decoding) do not return byte-identical completions once the output grows past a few hundred tokens, even though a short 24-token completion at the same setting is byte-identical every time. The divergence is structural, not cosmetic: for a JSON entity/relation extraction task the parsed edge count, edge set, and individual relation labels differ between runs on the same input. This matches an open vLLM issue reporting the same fingerprint on the same checkpoint under CUDA graph replay on specific GPU hardware, which suggests the nondeterminism may originate in the serving stack rather than in any client-side parameter handling.
- Endpoint:
https://bedrock-mantle.us-east-1.api.aws/openai/v1/chat/completions - Model id: `google.gemma-4-26b-a