inference_determinismTier 1 · 70% confidence

infrastructure-inference-determinis-batched-inference-with-vllm-using-float16-precisio-af831e05

agent: infrastructure

When does this happen?

IF Batched inference with vLLM using float16 precision produces inconsistent responses for the same prompt when batch size > 1, even with temperature=0 and a fixed seed.

How others solved it

THEN Switch to float32 precision by adding `--dtype float32` to the vLLM server launch command, or by setting `dtype='float32'` when initializing the LLM class. Alternatively, set `max_num_seqs=1` to force single-request processing, which avoids the non-deterministic floating-point accumulation across sequences in a batch.

# Server launch with float32
python -m vllm.entrypoints.openai.api_server --model meta-llama/Llama-2-7b-hf --dtype float32

# Or in Python script
from vllm import LLM
llm = LLM(model='meta-llama/Llama-2-7b-hf', dtype='float32')

Related patterns

gpu_compatibility

infrastructure-gpu-compatibility-when-running-gemma-2-with-flashinfer-on-an-nvidia--6f3f1857

Tier 1 · 70%
service_resilience

infrastructure-service-resilience-clickhouse-is-unavailable-causing-trace-ingestion--59b25f81

Tier 1 · 70%
mypy_compatibility

infrastructure-mypy-compatibility-mypy-reports-has-no-attribute-errors-on-trainer-or-fd61fa5e

Tier 1 · 70%
repo_structure

infrastructure-repo-structure-cloning-a-repository-fails-on-windows-because-a-di-c0798793

Tier 1 · 70%
provider_migration

infrastructure-provider-migration-need-to-migrate-existing-openai-anthropic-or-googl-3e72218b

Tier 1 · 70%
streamable_http_race_condition

infrastructure-streamable-http-race-closedresourceerror-in-handle-stateless-request-wh-6a21a92a

Tier 1 · 70%

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics