llama.cpp (fbe7dfa53caff0a7e830b676e6e949917a5c71b4) patch for miqu

a.patch

diff --git a/gguf-py/scripts/gguf-dump.py b/gguf-py/scripts/gguf-dump.py
index dbf89150..eeea947f 100755
--- a/gguf-py/scripts/gguf-dump.py
+++ b/gguf-py/scripts/gguf-dump.py
@@ -24,6 +24,56 @@ def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]:
         file_endian = host_endian
     return (host_endian, file_endian)
 
+from tqdm import tqdm
+import gguf.tensor_mapping as tm
+import re
+import torch
+LAYERS = 80
+INT_REGEX = re.compile(r'\.[0-9]+\.')
+def get_layer_idx(k: str): return int(INT_REGEX.search(k).group()[1:-1])
+def possibly_fix_key(k: str):
+    '''
+    [ 8192 32000]    output.weight  ->  lm_head.linear
+    [ 8192 32000]    token_embd.weight  ->  transformer.embd.wte
+    [8192]   output_norm.weight  ->  lm_head.ln
+    [8192]   blk.79.attn_norm.weight  ->  model.layers.layers.79.norm
+    [8192]   blk.79.ffn_norm.weight  ->  h.79.ln_2
+    [28672  8192]    blk.79.ffn_down.weight  ->  model.layers.layers.79.mlp.down_proj
+    [ 8192 28672]    blk.79.ffn_gate.weight  ->  model.layers.layers.79.mlp.gate_proj
+    [ 8192 28672]    blk.79.ffn_up.weight  ->  model.layers.layers.79.mlp.up_proj
+    [8192 1024]      blk.79.attn_k.weight  ->  model.layers.layers.79.self_attn.k_proj
+    [8192 8192]      blk.79.attn_output.weight  ->  model.layers.layers.79.self_attn.o_proj
+    [8192 8192]      blk.79.attn_q.weight  ->  model.layers.layers.79.self_attn.q_proj
+    [8192 1024]      blk.79.attn_v.weight  ->  model.layers.layers.79.self_attn.v_proj
+    '''
+    if k[0] == 'h' and k[-4:] == 'ln_2': # blk.79.ffn_norm.weight  ->  h.79.ln_2
+        k = f'model.layers.{get_layer_idx(k)}.post_attention_layernorm'
+    elif k[-4:] == 'norm':
+        k = f'model.layers.{get_layer_idx(k)}.input_layernorm'
+    
+    D = {'lm_head.linear': 'lm_head', 'transformer.embd.wte': 'model.embed_tokens', 'lm_head.ln': 'model.norm'}
+    if k in D: k = D[k]
+
+    k = k.replace('layers.layers', 'layers')
+    k += '.weight'
+
+    return k
+# see https://github.com/ggerganov/llama.cpp/blob/master/convert.py#L1182
+def possibly_permute(t: torch.Tensor, k: str):
+    if 'q_proj' in k or 'k_proj' in k:
+        H = 64 if 'q_proj' in k else 8
+        return t.reshape(H, t.shape[0] // H // 2, 2, *t.shape[1:]).swapaxes(1,2).reshape(t.shape)
+    return t
+def convert_to_hf(r: GGUFReader):
+    d = {v[1]:k for k,v in tm.get_tensor_name_map(tm.MODEL_ARCH.LLAMA, LAYERS).mapping.items()}
+    for rt in r.tensors: print(rt.shape, '\t', rt.name, ' -> ', d[rt.name[:-7]])
+    assert all(rt.data.dtype == np.float32 or rt.data.dtype == np.float16 for rt in r.tensors)
+    sd = {
+        possibly_fix_key(d[rt.name[:-7]]): torch.from_numpy(rt.data).half().view(*reversed(rt.shape.tolist()))
+        for rt in tqdm(r.tensors)
+    }
+    sd = {k: possibly_permute(v, k) for k,v in tqdm(sd.items()})
+    return sd
 
 # For more information about what field.parts and field.data represent,
 # please see the comments in the modify_gguf.py example.
@@ -103,10 +153,13 @@ def main() -> None:
     parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata")
     parser.add_argument("--json",       action="store_true", help="Produce JSON output")
     parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)")
+    parser.add_argument("--to-hf",      action="store_true")
     args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
     if not args.json:
         print(f'* Loading: {args.model}')
     reader = GGUFReader(args.model, 'r')
+    if args.to_hf:
+        return convert_to_hf(reader)
     if args.json:
         dump_metadata_json(reader, args)
     else:
@@ -114,4 +167,5 @@ def main() -> None:
 
 
 if __name__ == '__main__':
-    main()
+    c = main()
+    torch.save(c,"hf_pm2.pt")