gguf-py : add GGMLFileType

* convert-hf : use GGMLFileType

Author: compilade

convert-hf-to-gguf.py

@@ -239,10 +239,7 @@ def write_tensors(self):
                 data: np.ndarray = data  # type hint
                 n_dims = len(data.shape)
                 data_dtype = data.dtype
-
-                # if f32 desired, convert any float16 to float32
-                if self.ftype == 0 and data_dtype == np.float16:
-                    data = data.astype(np.float32)
+                data_qtype: gguf.GGMLQuantizationType | None = None
 
                 # when both are True, f32 should win
                 extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
@@ -254,20 +251,27 @@ def write_tensors(self):
                 # if f16 desired, convert any float32 2-dim weight tensors to float16
                 extra_f16 = extra_f16 or (name.endswith(".weight") and n_dims >= 2)
 
-                # when both extra_f32 and extra_f16 are False, convert to float32 by default
-                if self.ftype == 1 and data_dtype == np.float16 and (extra_f32 or not extra_f16):
-                    data = data.astype(np.float32)
+                if self.ftype != gguf.GGMLFileType.ALL_F32 and extra_f16 and not extra_f32:
+                    if self.ftype == gguf.GGMLFileType.MOSTLY_F16:
+                        if data_dtype != np.float16:
+                            data = data.astype(np.float16)
+                        data_qtype = gguf.GGMLQuantizationType.F16
+                    # TODO: add more types (like BF16) here
+
+                else:  # by default, convert to float32
+                    if data_dtype != np.float32:
+                        data = data.astype(np.float32)
+                    data_qtype = gguf.GGMLQuantizationType.F32
 
-                if self.ftype == 1 and data_dtype == np.float32 and extra_f16 and not extra_f32:
-                    data = data.astype(np.float16)
+                assert data_qtype is not None
 
                 # reverse shape to make it similar to the internal ggml dimension order
                 shape_str = f"{{{', '.join(str(n) for n in reversed(data.shape))}}}"
 
                 # n_dims is implicit in the shape
-                logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data.dtype}, shape = {shape_str}")
+                logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
 
-                self.gguf_writer.add_tensor(new_name, data)
+                self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype)
 
     def write(self):
         self.write_tensors()
@@ -2472,9 +2476,9 @@ def main() -> None:
         logger.error(f'Error: {args.model} is not a directory')
         sys.exit(1)
 
-    ftype_map = {
-        "f32": gguf.GGMLQuantizationType.F32,
-        "f16": gguf.GGMLQuantizationType.F16,
+    ftype_map: dict[str, gguf.GGMLFileType] = {
+        "f32": gguf.GGMLFileType.ALL_F32,
+        "f16": gguf.GGMLFileType.MOSTLY_F16,
     }
 
     if args.outfile is not None:

gguf-py/gguf/constants.py

@@ -820,6 +820,42 @@ class GGMLQuantizationType(IntEnum):
     BF16    = 30
 
 
+class GGMLFileType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1   # except 1d tensors
+    MOSTLY_Q4_0          = 2   # except 1d tensors
+    MOSTLY_Q4_1          = 3   # except 1d tensors
+    MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
+    # MOSTLY_Q4_2        = 5   # support has been removed
+    # MOSTLY_Q4_3        = 6   # support has been removed
+    MOSTLY_Q8_0          = 7   # except 1d tensors
+    MOSTLY_Q5_0          = 8   # except 1d tensors
+    MOSTLY_Q5_1          = 9   # except 1d tensors
+    MOSTLY_Q2_K          = 10  # except 1d tensors
+    MOSTLY_Q3_K_S        = 11  # except 1d tensors
+    MOSTLY_Q3_K_M        = 12  # except 1d tensors
+    MOSTLY_Q3_K_L        = 13  # except 1d tensors
+    MOSTLY_Q4_K_S        = 14  # except 1d tensors
+    MOSTLY_Q4_K_M        = 15  # except 1d tensors
+    MOSTLY_Q5_K_S        = 16  # except 1d tensors
+    MOSTLY_Q5_K_M        = 17  # except 1d tensors
+    MOSTLY_Q6_K          = 18  # except 1d tensors
+    MOSTLY_IQ2_XXS       = 19  # except 1d tensors
+    MOSTLY_IQ2_XS        = 20  # except 1d tensors
+    MOSTLY_Q2_K_S        = 21  # except 1d tensors
+    MOSTLY_IQ3_XS        = 22  # except 1d tensors
+    MOSTLY_IQ3_XXS       = 23  # except 1d tensors
+    MOSTLY_IQ1_S         = 24  # except 1d tensors
+    MOSTLY_IQ4_NL        = 25  # except 1d tensors
+    MOSTLY_IQ3_S         = 26  # except 1d tensors
+    MOSTLY_IQ3_M         = 27  # except 1d tensors
+    MOSTLY_IQ2_S         = 28  # except 1d tensors
+    MOSTLY_IQ2_M         = 29  # except 1d tensors
+    MOSTLY_IQ4_XS        = 30  # except 1d tensors
+    MOSTLY_IQ1_M         = 31  # except 1d tensors
+    MOSTLY_BF16          = 32  # except 1d tensors
+
+
 class GGUFEndian(IntEnum):
     LITTLE = 0
     BIG = 1