From 1b585b0d40c991cad9eb1bf3b38a7856a186840c Mon Sep 17 00:00:00 2001 From: "Chen, Zhentao" Date: Mon, 8 Jan 2024 15:53:57 +0800 Subject: [PATCH] set fp8 default as e5m2 (#9859) --- python/llm/src/bigdl/llm/ggml/quantize.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/llm/src/bigdl/llm/ggml/quantize.py b/python/llm/src/bigdl/llm/ggml/quantize.py index 185d31a7..11e04ab8 100644 --- a/python/llm/src/bigdl/llm/ggml/quantize.py +++ b/python/llm/src/bigdl/llm/ggml/quantize.py @@ -38,7 +38,7 @@ ggml_tensor_qtype = {"sym_int4": 2, # q4_0 in ggml "mixed_fp4": 17, # Mixture of Formats Quantization 4 bits "mixed_fp8": 18, # Mixture of Formats Quantization 8 bits "fp8_e5m2": 19, # fp8 in e5m2 format - "fp8": 15, # fp8 in e4m3 format + "fp8": 19, # fp8 in e5m2 format "bf16": 20} _llama_quantize_type = {"q4_0": 2,