Merge 6847b3d86692438b63a831f7649705c18d7d333b into 9b4e9788e4a3a731f7567338ed15d3ec549ce03b

README.md

@@ -246,6 +246,11 @@ cd inference
 python fp8_cast_bf16.py --input-fp8-hf-path /path/to/fp8_weights --output-bf16-hf-path /path/to/bf16_weights
 ```
 
+> [!TIP]
+> The conversion script now supports optional arguments for broader portability:
+> - `--device`: `auto` (default), `cuda`, or `cpu`. When set to `auto`, the script prefers CUDA if available, otherwise falls back to CPU. Use `cpu` to force a CPU-only path on systems without GPUs.
+> - `--block-size`: Block size used during quantization/dequantization (default `128`). This should match the model’s tiling settings at export time.
+
 > [!NOTE]
 > Hugging Face's Transformers has not been directly supported yet.
 

inference/fp8_cast_bf16.py

@@ -2,14 +2,47 @@ import os
 import json
 from argparse import ArgumentParser
 from glob import glob
+from typing import Dict
 from tqdm import tqdm
 
 import torch
 from safetensors.torch import load_file, save_file
 
-from kernel import weight_dequant
+# Use the optimized Triton GPU kernel when available; add a CPU fallback below
+from kernel import weight_dequant as weight_dequant_gpu
 
-def main(fp8_path, bf16_path):
+def _infer_device(requested_device: str) -> str:
+    """
+    Decide which device to use for tensor I/O and dequantization.
+
+    - "auto": prefer CUDA if available, otherwise CPU
+    - "cuda": use CUDA when available, else fall back to CPU
+    - "cpu": force CPU
+    """
+    if requested_device == "cuda":
+        return "cuda" if torch.cuda.is_available() else "cpu"
+    if requested_device == "cpu":
+        return "cpu"
+    return "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def _weight_dequant_cpu(x: torch.Tensor, s: torch.Tensor, block_size: int = 128) -> torch.Tensor:
+    """
+    CPU fallback for FP8 weight dequantization using pure PyTorch.
+
+    Expands per-block scales to a full resolution map and rescales FP8 weights.
+    Prioritizes correctness and portability; may be memory intensive for huge tensors.
+    """
+    assert x.dim() == 2 and s.dim() == 2, "Input tensors must have 2 dimensions"
+    M, N = x.shape
+    m_blocks, n_blocks = s.shape
+    assert M % block_size == 0 and N % block_size == 0, "Weight dims must be multiples of block_size"
+    assert m_blocks == M // block_size and n_blocks == N // block_size, "Scale shape must match weight tiling"
+    s_full = s.repeat_interleave(block_size, dim=0).repeat_interleave(block_size, dim=1)
+    y = x.to(torch.float32) * s_full
+    return y.to(torch.get_default_dtype())
+
+def convert_fp8_to_bf16(fp8_path: str, bf16_path: str, device: str = "auto", block_size: int = 128):
     """
     Converts FP8 weights to BF16 and saves the converted weights.
 
@@ -20,6 +53,8 @@ def main(fp8_path, bf16_path):
     Args:
     fp8_path (str): The path to the directory containing the FP8 weights and model index file.
     bf16_path (str): The path to the directory where the converted BF16 weights will be saved.
+    device (str): One of {"auto", "cuda", "cpu"}. Controls load/compute device.
+    block_size (int): Block size used by quantization/dequantization. Typically 128.
 
     Raises:
     KeyError: If a required scale_inv tensor is missing for a weight.
@@ -31,13 +66,14 @@ def main(fp8_path, bf16_path):
     """
     torch.set_default_dtype(torch.bfloat16)
     os.makedirs(bf16_path, exist_ok=True)
+    effective_device = _infer_device(device)
     model_index_file = os.path.join(fp8_path, "model.safetensors.index.json")
     with open(model_index_file, "r") as f:
         model_index = json.load(f)
     weight_map = model_index["weight_map"]
     
     # Cache for loaded safetensor files
-    loaded_files = {}
+    loaded_files: Dict[str, Dict[str, torch.Tensor]] = {}
     fp8_weight_names = []
 
     # Helper function to get tensor from the correct file
@@ -57,14 +93,14 @@ def main(fp8_path, bf16_path):
         file_name = weight_map[tensor_name]
         if file_name not in loaded_files:
             file_path = os.path.join(fp8_path, file_name)
-            loaded_files[file_name] = load_file(file_path, device="cuda")
+            loaded_files[file_name] = load_file(file_path, device=effective_device)
         return loaded_files[file_name][tensor_name]
 
     safetensor_files = list(glob(os.path.join(fp8_path, "*.safetensors")))
     safetensor_files.sort()
     for safetensor_file in tqdm(safetensor_files):
         file_name = os.path.basename(safetensor_file)
-        current_state_dict = load_file(safetensor_file, device="cuda")
+        current_state_dict = load_file(safetensor_file, device=effective_device)
         loaded_files[file_name] = current_state_dict
         
         new_state_dict = {}
@@ -77,7 +113,10 @@ def main(fp8_path, bf16_path):
                     # Get scale_inv from the correct file
                     scale_inv = get_tensor(scale_inv_name)
                     fp8_weight_names.append(weight_name)
-                    new_state_dict[weight_name] = weight_dequant(weight, scale_inv)
+                    if effective_device == "cuda":
+                        new_state_dict[weight_name] = weight_dequant_gpu(weight, scale_inv, block_size=block_size)
+                    else:
+                        new_state_dict[weight_name] = _weight_dequant_cpu(weight, scale_inv, block_size=block_size)
                 except KeyError:
                     print(f"Warning: Missing scale_inv tensor for {weight_name}, skipping conversion")
                     new_state_dict[weight_name] = weight
@@ -91,7 +130,8 @@ def main(fp8_path, bf16_path):
         if len(loaded_files) > 2:
             oldest_file = next(iter(loaded_files))
             del loaded_files[oldest_file]
-            torch.cuda.empty_cache()
+            if effective_device == "cuda":
+                torch.cuda.empty_cache()
     
     # Update model index
     new_model_index_file = os.path.join(bf16_path, "model.safetensors.index.json")
@@ -107,6 +147,10 @@ if __name__ == "__main__":
     parser = ArgumentParser()
     parser.add_argument("--input-fp8-hf-path", type=str, required=True)
     parser.add_argument("--output-bf16-hf-path", type=str, required=True)
+    parser.add_argument("--device", type=str, default="auto", choices=["auto", "cuda", "cpu"],
+                        help="Select compute device: 'auto' prefers CUDA if available; otherwise CPU.")
+    parser.add_argument("--block-size", type=int, default=128,
+                        help="Block size used during quantization/dequantization. Typically 128.")
     args = parser.parse_args()
-    main(args.input_fp8_hf_path, args.output_bf16_hf_path)
+    convert_fp8_to_bf16(args.input_fp8_hf_path, args.output_bf16_hf_path, device=args.device, block_size=args.block_size)
     

inference/tests/test_fp8_cast_bf16.py

@@ -0,0 +1,74 @@
+import os
+import json
+import tempfile
+from typing import Tuple
+
+import torch
+from safetensors.torch import save_file, load_file
+
+# Import the conversion API we expose for programmatic use
+from inference.fp8_cast_bf16 import convert_fp8_to_bf16
+
+
+def _make_block_scale(shape_blocks: Tuple[int, int], value: float, device: str) -> torch.Tensor:
+    """
+    Create a per-block scale tensor of shape (M_blocks, N_blocks) filled with a constant.
+    """
+    return torch.full(shape_blocks, value, dtype=torch.float32, device=device).contiguous()
+
+
+def test_convert_fp8_to_bf16_cpu_roundtrip_small_matrix():
+    """
+    Validate CPU fallback by constructing a tiny FP8 weight with known block scales,
+    converting to BF16, and checking the recovered values.
+    """
+    if not hasattr(torch, "float8_e4m3fn"):
+        # Skip if PyTorch build lacks float8 support
+        return
+
+    device = "cpu"
+    block_size = 2
+    M, N = 4, 4
+    # Choose a uniform block scale that is easy to reason about
+    scale_value = 0.5  # multiplicative factor used during dequant
+    # Construct the target dequantized weights (what we want to recover)
+    y_true = torch.tensor(
+        [[1.0, 2.0, 3.0, 4.0],
+         [5.0, 6.0, 7.0, 8.0],
+         [2.0, 4.0, 6.0, 8.0],
+         [1.5, 2.5, 3.5, 4.5]],
+        dtype=torch.float32,
+        device=device,
+    )
+    # Create the per-block scale tensor: (M // block_size, N // block_size)
+    s = _make_block_scale((M // block_size, N // block_size), scale_value, device)
+    # Expand s to full resolution for constructing FP8 quantized weights
+    s_full = s.repeat_interleave(block_size, dim=0).repeat_interleave(block_size, dim=1)
+    # Build FP8 weights such that dequant (x * s_full) recovers y_true
+    x_fp32 = (y_true / scale_value).contiguous()
+    x_fp8 = x_fp32.to(torch.float8_e4m3fn)
+
+    with tempfile.TemporaryDirectory() as tmp:
+        fp8_dir = os.path.join(tmp, "fp8")
+        bf16_dir = os.path.join(tmp, "bf16")
+        os.makedirs(fp8_dir, exist_ok=True)
+        os.makedirs(bf16_dir, exist_ok=True)
+
+        # Create minimal safetensors shard and index
+        shard = {"layer.weight": x_fp8, "layer.weight_scale_inv": s}
+        shard_name = "model-00001-of-00001.safetensors"
+        save_file(shard, os.path.join(fp8_dir, shard_name))
+
+        index = {"metadata": {}, "weight_map": {"layer.weight": shard_name, "layer.weight_scale_inv": shard_name}}
+        with open(os.path.join(fp8_dir, "model.safetensors.index.json"), "w") as f:
+            json.dump(index, f)
+
+        # Run conversion using CPU path and a small block size
+        convert_fp8_to_bf16(fp8_dir, bf16_dir, device="cpu", block_size=block_size)
+
+        # Load converted weights and verify they match the expected y_true (within tolerance)
+        out_shard = load_file(os.path.join(bf16_dir, shard_name), device=device)
+        y = out_shard["layer.weight"].to(torch.float32)
+        assert torch.allclose(y, y_true, atol=1e-2, rtol=1e-2)
+
+