LLM: support starcoder in llm-cli (#8377)

* support starcoder in cli

* small fix

python/llm/src/bigdl/llm/cli/llm-cli

@@ -53,6 +53,12 @@ function gptneox {
   eval "$command"
 }
 
+function starcoder {
+  command="$lib_dir/main-starcoder_$avx_flag -t $threads -n $n_predict ${filteredArguments[*]}"
+  echo "$command"
+  eval "$command"
+}
+
 # Remove model_family/x parameter
 filteredArguments=()
 while [[ $# -gt 0 ]]; do
@@ -91,6 +97,8 @@ elif [[ "$model_family" == "bloom" ]]; then
   bloom
 elif [[ "$model_family" == "gptneox" ]]; then
   gptneox
+elif [[ "$model_family" == "starcoder" ]]; then
+  starcoder
 else
   echo "Invalid model_family: $model_family"
   display_help

python/llm/src/bigdl/llm/cli/llm-cli.ps1

@@ -40,6 +40,13 @@ function gptneox
     Invoke-Expression $command
 }
 
+function starcoder
+{
+    $command = "$lib_dir/main-starcoder.exe -t $threads -n $n_predict $filteredArguments"
+    Write-Host "$command"
+    Invoke-Expression $command
+}
+
 # Remove model_family/x parameter
 $filteredArguments = @()
 for ($i = 0; $i -lt $args.Length; $i++) {
@@ -79,6 +86,9 @@ switch ($model_family)
     "gptneox" {
         gptneox
     }
+    "starcoder" {
+        starcoder
+    }
     default {
         Write-Host "Invalid model_family: $model_family"
         Display-Help

python/llm/src/bigdl/llm/models.py

@@ -22,3 +22,4 @@
 from bigdl.llm.ggml.model.llama import Llama
 from bigdl.llm.ggml.model.gptneox import Gptneox
 from bigdl.llm.ggml.model.bloom import Bloom
+from bigdl.llm.ggml.model.starcoder import Starcoder

python/llm/src/bigdl/llm/utils/convert_util.py

@@ -1427,7 +1427,7 @@ def _convert_starcoder_hf_to_ggml(model_path, outfile_dir, outtype):
     model = AutoModelForCausalLM.from_pretrained(model_path, config=config,
                                                  torch_dtype=torch.float16
                                                  if outtype == "f16" else torch.float32,
-                                                 low_cpu_mem_usage=True,
+                                                 # low_cpu_mem_usage=True,
                                                  trust_remote_code=True,
                                                  offload_state_dict=True)