add script/convert_fp8_scale_to_bf16.py

script/convert_fp8_scale_to_bf16.py

@@ -0,0 +1,283 @@
+#!/usr/bin/env python
+import argparse
+import json
+import math
+import os
+import struct
+from collections import Counter
+from pathlib import Path
+
+import torch
+from safetensors import safe_open
+
+
+FLOAT_DTYPES = {
+    "BF16",
+    "F16",
+    "F32",
+    "F64",
+    "F8_E4M3",
+    "F8_E4M3FN",
+    "F8_E5M2",
+}
+
+FP8_DTYPES = {
+    "F8_E4M3",
+    "F8_E4M3FN",
+    "F8_E5M2",
+}
+
+DTYPE_SIZES = {
+    "BOOL": 1,
+    "U8": 1,
+    "I8": 1,
+    "F8_E4M3": 1,
+    "F8_E4M3FN": 1,
+    "F8_E5M2": 1,
+    "U16": 2,
+    "I16": 2,
+    "F16": 2,
+    "BF16": 2,
+    "U32": 4,
+    "I32": 4,
+    "F32": 4,
+    "U64": 8,
+    "I64": 8,
+    "F64": 8,
+}
+
+
+def read_safetensors_header(path: Path):
+    with path.open("rb") as f:
+        header_len = struct.unpack("<Q", f.read(8))[0]
+        header = f.read(header_len).decode("utf-8").rstrip()
+    return json.loads(header)
+
+
+def numel(shape):
+    return math.prod(shape) if shape else 1
+
+
+def scale_key_for_weight(name: str):
+    if name.endswith(".weight"):
+        return name[:-len(".weight")] + ".weight_scale"
+    if name.endswith("weight"):
+        return name + "_scale"
+    return None
+
+
+def tensor_nbytes(dtype: str, shape):
+    return numel(shape) * DTYPE_SIZES[dtype]
+
+
+def build_output_plan(header):
+    entries = {k: v for k, v in header.items() if k != "__metadata__"}
+    paired_scale_keys = set()
+    plan = []
+
+    for name, info in entries.items():
+        scale_key = scale_key_for_weight(name)
+        if info["dtype"] in FP8_DTYPES and scale_key in entries:
+            paired_scale_keys.add(scale_key)
+
+    for name, info in entries.items():
+        if name in paired_scale_keys:
+            continue
+
+        dtype = info["dtype"]
+        shape = info["shape"]
+        scale_key = scale_key_for_weight(name)
+
+        if dtype in FP8_DTYPES and scale_key in entries:
+            scale_info = entries[scale_key]
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": "BF16",
+                    "shape": shape,
+                    "mode": "fp8_scaled_weight",
+                    "scale_key": scale_key,
+                }
+            )
+            continue
+
+        if dtype in FLOAT_DTYPES:
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": "BF16",
+                    "shape": shape,
+                    "mode": "float_to_bf16",
+                }
+            )
+        else:
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": dtype,
+                    "shape": shape,
+                    "mode": "copy",
+                }
+            )
+
+    metadata = dict(header.get("__metadata__", {}) or {})
+    metadata["format"] = "pt"
+    metadata["conversion"] = "fp8_weight_scale_to_bf16"
+
+    output_header = {"__metadata__": metadata}
+    offset = 0
+    for item in plan:
+        size = tensor_nbytes(item["output_dtype"], item["shape"])
+        output_header[item["name"]] = {
+            "dtype": item["output_dtype"],
+            "shape": item["shape"],
+            "data_offsets": [offset, offset + size],
+        }
+        offset += size
+
+    return plan, output_header, offset
+
+
+def write_tensor_bytes(out, tensor):
+    tensor = tensor.detach().cpu().contiguous()
+    if tensor.numel() == 0:
+        return
+    if tensor.dtype == torch.bfloat16:
+        tensor.view(torch.uint16).numpy().tofile(out)
+    elif tensor.dtype in (getattr(torch, "float8_e4m3fn", None), getattr(torch, "float8_e5m2", None)):
+        tensor.view(torch.uint8).numpy().tofile(out)
+    else:
+        tensor.numpy().tofile(out)
+
+
+def scale_view_for_chunk(scale, chunk, first_dim_start=0, first_dim_end=None):
+    scale = scale.to(torch.float32)
+
+    if scale.numel() == 1:
+        return scale.reshape((1,) * chunk.ndim)
+
+    if chunk.ndim > 0 and scale.ndim == 1:
+        if first_dim_end is not None and scale.shape[0] >= first_dim_end:
+            scale = scale[first_dim_start:first_dim_end]
+        if scale.shape[0] == chunk.shape[0]:
+            return scale.reshape((scale.shape[0],) + (1,) * (chunk.ndim - 1))
+
+    return scale
+
+
+def write_scaled_fp8_weight(out, weight, scale, chunk_rows):
+    if weight.ndim == 0:
+        result = weight.to(torch.float32) * scale_view_for_chunk(scale, weight)
+        write_tensor_bytes(out, result.to(torch.bfloat16))
+        return
+
+    rows = weight.shape[0]
+    for start in range(0, rows, chunk_rows):
+        end = min(start + chunk_rows, rows)
+        chunk = weight[start:end].to(torch.float32)
+        scale_view = scale_view_for_chunk(scale, chunk, start, end)
+        result = chunk * scale_view
+        write_tensor_bytes(out, result.to(torch.bfloat16))
+
+
+def write_float_as_bf16(out, tensor, chunk_rows):
+    if tensor.dtype == torch.bfloat16:
+        write_tensor_bytes(out, tensor)
+        return
+
+    if tensor.ndim == 0:
+        write_tensor_bytes(out, tensor.to(torch.bfloat16))
+        return
+
+    rows = tensor.shape[0]
+    for start in range(0, rows, chunk_rows):
+        end = min(start + chunk_rows, rows)
+        write_tensor_bytes(out, tensor[start:end].to(torch.bfloat16))
+
+
+def convert(input_path: Path, output_path: Path, chunk_rows: int, dry_run: bool):
+    header = read_safetensors_header(input_path)
+    plan, output_header, data_size = build_output_plan(header)
+
+    source_counts = Counter(item["source_dtype"] for item in plan)
+    output_counts = Counter(item["output_dtype"] for item in plan)
+    scaled_count = sum(item["mode"] == "fp8_scaled_weight" for item in plan)
+    dropped_scales = sum(item["mode"] == "fp8_scaled_weight" for item in plan)
+    header_bytes = json.dumps(output_header, separators=(",", ":")).encode("utf-8")
+    expected_size = 8 + len(header_bytes) + data_size
+
+    print(f"input:  {input_path}")
+    print(f"output: {output_path}")
+    print(f"tensors written: {len(plan)}")
+    print(f"scaled fp8 weights dequantized: {scaled_count}")
+    print(f"weight_scale tensors dropped: {dropped_scales}")
+    print(f"source dtypes: {dict(sorted(source_counts.items()))}")
+    print(f"output dtypes: {dict(sorted(output_counts.items()))}")
+    print(f"expected output size: {expected_size / (1024 ** 3):.2f} GiB")
+
+    if dry_run:
+        return
+
+    if output_path.exists():
+        raise FileExistsError(f"{output_path} already exists; pass --overwrite to replace it")
+
+    tmp_path = output_path.with_suffix(output_path.suffix + ".tmp")
+    if tmp_path.exists():
+        raise FileExistsError(f"{tmp_path} already exists; remove it or choose another output")
+
+    with safe_open(str(input_path), framework="pt", device="cpu") as sf, tmp_path.open("wb") as out:
+        out.write(struct.pack("<Q", len(header_bytes)))
+        out.write(header_bytes)
+
+        for index, item in enumerate(plan, 1):
+            name = item["name"]
+            print(f"[{index:04d}/{len(plan):04d}] {name} -> {item['output_dtype']}")
+
+            tensor = sf.get_tensor(name)
+            if item["mode"] == "fp8_scaled_weight":
+                scale = sf.get_tensor(item["scale_key"])
+                write_scaled_fp8_weight(out, tensor, scale, chunk_rows)
+            elif item["mode"] == "float_to_bf16":
+                write_float_as_bf16(out, tensor, chunk_rows)
+            else:
+                write_tensor_bytes(out, tensor)
+
+        actual_size = out.tell()
+
+    if actual_size != expected_size:
+        tmp_path.unlink(missing_ok=True)
+        raise RuntimeError(f"wrote {actual_size} bytes, expected {expected_size} bytes")
+
+    tmp_path.replace(output_path)
+    print("done")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Convert an fp8 safetensors checkpoint with weight_scale tensors to bf16."
+    )
+    parser.add_argument("--input", default="ideogram4_fp8.safetensors", type=Path)
+    parser.add_argument("--output", default="ideogram4_bf16.safetensors", type=Path)
+    parser.add_argument("--chunk-rows", default=1024, type=int)
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--overwrite", action="store_true")
+    args = parser.parse_args()
+
+    input_path = args.input.resolve()
+    output_path = args.output.resolve()
+
+    if args.chunk_rows < 1:
+        raise ValueError("--chunk-rows must be >= 1")
+    if not input_path.exists():
+        raise FileNotFoundError(input_path)
+    if args.overwrite and output_path.exists():
+        output_path.unlink()
+
+    convert(input_path, output_path, args.chunk_rows, args.dry_run)
+
+
+if __name__ == "__main__":
+    main()