upload host_weight folder

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+host_weight/Whead_fp.csv filter=lfs diff=lfs merge=lfs -text
+host_weight/emb.csv filter=lfs diff=lfs merge=lfs -text

host_weight/Makefile

@@ -0,0 +1,21 @@
+PHONY: 
+
+# path to original model weights
+MODEL_DIR = ../../quant/hf_weight
+
+# file names
+EMB = emb
+WHEAD = Whead_fp
+
+clean:
+	rm -f *.csv
+	rm -f *.bin
+
+gen_csv:
+	rm -f *.csv
+	python gen_csv.py --model_dir $(MODEL_DIR) --emb $(EMB) --whead $(WHEAD)
+
+# wait for several seconds
+gen_bin:
+	rm -f *.bin
+	python gen_bin.py --emb $(EMB) --whead $(WHEAD)

host_weight/README.md

@@ -0,0 +1,10 @@
+### 產生 embed 和 lm_head
+
+### embed
+流程: .safetensors -> emb.csv -> emb.bin -> emb_elem.bin & emb_scale.bin
+
+最後一個步驟要利用 host_code 當中的 mxq_emb 來產生 emb_elem.bin 和 emb_scale.bin   
+但目前 demo 所使用的 cpp 檔為 llm_acc_host.cpp 暫時用不到 emb_elem.bin 和 emb_scale.bin
+
+### lm_head
+流程: .safetensors -> Whead_fp.csv -> Whead_fp.bin

host_weight/Whead_fp.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89ad3853f57938a4634a960a2f0b0b4f1887699aefc338d0317fd0b8333d457f
+size 524288000

host_weight/Whead_fp.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a2fa11f326f0c477af049ab46bacba72470e40215066ff42ab7a96f6e8f2a26
+size 3342321439

host_weight/emb.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4f79a4374a649dfd30b70325c80bf824d12ca9b826a278706d4392a73e7f597
+size 524288000

host_weight/emb.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f94761a5565853c25ebda677de48fdc52a69a0acc82922a3c90979dd6f2264ed
+size 3342488234

host_weight/gen_bin.py

@@ -0,0 +1,31 @@
+import numpy as np
+import struct
+import argparse
+
+parser = argparse.ArgumentParser(description="Generate binary files from CSV embeddings and weights.")
+parser.add_argument("--emb", type=str, default="emb", help="Path to the embedding CSV file.")
+parser.add_argument("--whead", type=str, default="Whead_fp", help="Path to the Whead CSV file.")
+args = parser.parse_args()
+
+if __name__ == '__main__':
+    emb_path = args.emb + ".csv"
+    head_path = args.whead + ".csv"
+
+    emb = np.loadtxt(emb_path, delimiter=',')
+    Whead = np.loadtxt(head_path, delimiter=',')
+
+    print(emb.shape)
+    print(Whead.shape)
+    
+    emb = emb.flatten()
+    Whead = Whead.flatten()
+
+    with open(args.emb + ".bin", 'wb') as f:
+        for data in emb:
+            binary_data = struct.pack('f', data)
+            f.write(binary_data)
+
+    with open(args.whead + ".bin", 'wb') as f:
+        for data in Whead:
+            binary_data = struct.pack('f', data)
+            f.write(binary_data)

host_weight/gen_csv.py

@@ -0,0 +1,30 @@
+from transformers import AutoModelForCausalLM
+import argparse
+import numpy as np
+import os
+
+def write_csv_chunked(path, array, chunk_rows=1024):
+    with open(path, "w") as f:
+        n_rows = array.shape[0]
+        for start in range(0, n_rows, chunk_rows):
+            end = min(start + chunk_rows, n_rows)
+            np.savetxt(f, array[start:end], delimiter=",")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Generate CSV file with model information.")
+    parser.add_argument("--model_dir", type=str, default="../../quant/hf_weight", help="Directory to save the model files.")
+    parser.add_argument("--emb", type=str, default="emb", help="Path to save the embedding CSV file.")
+    parser.add_argument("--whead", type=str, default="Whead_fp", help="Path to save the Whead CSV file.")
+    args = parser.parse_args()
+    model = AutoModelForCausalLM.from_pretrained(args.model_dir)
+
+    emb = model.model.embed_tokens.weight.detach().cpu().numpy()  # (V, H)
+    whead = model.lm_head.weight.detach().cpu().numpy()
+
+    emb_csv = args.emb + ".csv"
+    write_csv_chunked(emb_csv, emb)
+    print(f"Saved embeddings to {emb_csv}.")
+    whead_csv = args.whead + ".csv"
+    write_csv_chunked(whead_csv, whead)
+    print(f"Saved Whead to {whead_csv}.")
+    print("Conversion to binary completed.")