diff --git a/clusters/k3s-dgx/nim-service/qwen.yaml b/clusters/k3s-dgx/nim-service/qwen.yaml index f08ef2d..082a330 100644 --- a/clusters/k3s-dgx/nim-service/qwen.yaml +++ b/clusters/k3s-dgx/nim-service/qwen.yaml @@ -65,7 +65,7 @@ spec: - "254" # ↑ Allow more concurrent sequences - --enable-prefix-caching - --dtype - - "float8" # Explicit FP8 encoding + - "float" # Explicit FP8 encoding - --quantization - "fbgemm-fp8" # Explicit quantization backend - --sv2-transformer-bindings # Skip SA checks (small speedup)