Merge branch 'phi2' of https://github.com/intel/neural-speed into phi2

intel · Jan 25, 2024 · 1a86d12 · 1a86d12
2 parents 3702d1a + 105fab6
commit 1a86d12
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 7 deletions.
diff --git a/requirements.txt b/requirements.txt
@@ -13,5 +13,4 @@ tiktoken
 py-cpuinfo
 gguf
 cmake
-auto-gptq
 setuptools>=61
diff --git a/scripts/cal_diff.py b/scripts/cal_diff.py
@@ -35,10 +35,10 @@ def cmpData(numa, numb):
     args = parser.parse_args()
 
     woq_configs = {
-            "fp32": {"use_cache":True, "use_quant":False},
-        # "ggml_int4": {"compute_dtype":"int8", "weight_dtype":"int4", "use_cache":True, "use_ggml":True},
-        "jblas_int4": {"compute_dtype":"int8", "weight_dtype":"int4", "use_cache":True},
-        # "jblas_int8": {"compute_dtype":"bf16", "weight_dtype":"int8", "use_cache":True},
+            "fp32": {"use_quant":False},
+        # "ggml_int4": {"compute_dtype":"int8", "weight_dtype":"int4", "use_ggml":True},
+        "jblas_int4": {"compute_dtype":"int8", "weight_dtype":"int4"},
+        # "jblas_int8": {"compute_dtype":"bf16", "weight_dtype":"int8"},
 
     }
     prompt = "What is the meaning of life?"

diff --git a/tests/model-test/cpp_graph_inference.sh b/tests/model-test/cpp_graph_inference.sh
@@ -260,9 +260,9 @@ function main() {
         infer_cmd="./build/bin/run_whisper"
         precision_list+=("q4_0")
     elif [[ "${model}" == "phi2" ]]; then
-        quant_script="./build/bin/quant_phi2"
+        quant_script="./build/bin/quant_phi"
         convert_script="${convert_script}/convert_phi.py"
-        infer_cmd="./build/bin/run_phi2"
+        infer_cmd="./build/bin/run_phi"
     else
         echo "Error: Unexpedted model: $model" 1>&2
         exit 1