# mypy: allow-untyped-defs import json import logging import math from pathlib import Path from typing import Any import torch from torch.distributed.checkpoint._hf_utils import _metadata_fn from torch.distributed.checkpoint.metadata import TensorStorageMetadata from torch.distributed.checkpoint.planner import LoadPlanner, ReadItem from .hf_storage import HuggingFaceStorageReader logger: logging.Logger = logging.getLogger(__name__) __all__ = ["QuantizedHuggingFaceStorageReader"] class QuantizedHuggingFaceStorageReader(HuggingFaceStorageReader): """ Extension of HuggingFaceStorageReader that handles quantized tensors. Checkpoint should have the full tensor in a SafeTensor file. The quantized tensor should not be sharded across multiple files. This reader handles the dequantization of tensors during the read process, converting them from quantized blocks to full dequantized tensors before copying to the target tensor. """ def __init__( self, path: str, thread_count: int = 1, target_dtype: torch.dtype = torch.float32, block_size: int = 128, ): """ Initialize the HuggingFace storage reader to load quantized checkpoints Args: path: directory where the checkpoint will be read from. thread_count: Number of threads to use to read distributed checkpoint. Defaults to 1. target_dtype: Target dtype for dequantized tensor. Defaults to torch.float32. block_size: Fixed block size for dequantization. Defaults to 128. """ super().__init__(path=path, thread_count=thread_count) self.target_dtype: torch.dtype = target_dtype self.block_size: int = block_size self._weight_scale_mapping: dict[str, str] = {} # Track which file contains each tensor self._weight_map: dict[str, str] = {} # Cache for full tensor shapes (fqn -> shape) self._tensor_full_shapes: dict[str, torch.Size] = {} def read_metadata(self) -> Any: metadata = super().read_metadata() # Load quantization metadata first. self._load_quantization_metadata() # Build a cache of FQN -> full tensor shape, correcting for quantized tensors. for fqn, tensor_metadata in metadata.state_dict_metadata.items(): # Only process TensorStorageMetadata which has size attribute. if isinstance(tensor_metadata, TensorStorageMetadata): # Check if this is a MXFP4 quantized tensor that needs shape correction. if fqn.endswith("_blocks"): # Save the quantized tensor shapes for lookup when dequantization. self._tensor_full_shapes[fqn + "_quantized"] = tensor_metadata.size *prefix_shape, G, B = tensor_metadata.size dequantized_size = torch.Size([*prefix_shape, G * B * 2]) # Update the metadata with the size after dequantization. # Metadata used by planner to slice state dict. tensor_metadata.size = dequantized_size self._tensor_full_shapes[fqn] = dequantized_size else: self._tensor_full_shapes[fqn] = tensor_metadata.size return metadata def _load_quantization_metadata(self): """Load quantization metadata from the checkpoint.""" checkpoint_path = Path(self.path) # Load weight mapping from index file index_file = checkpoint_path / _metadata_fn with open(index_file) as f: index_data = json.load(f) weight_map = index_data.get("weight_map", {}) self._build_weight_scale_mapping(weight_map) def _build_weight_scale_mapping(self, weight_map: dict[str, str]): """Analyze and build weight-scale tensor pairs from weight mapping.""" # Store the complete weight map for file location lookups. self._weight_map = weight_map for tensor_name in weight_map: if tensor_name.endswith(".weight_scale_inv"): weight_name = tensor_name.replace(".weight_scale_inv", ".weight") if weight_name in weight_map: self._weight_scale_mapping[weight_name] = tensor_name # Handle MXFP4 format: _blocks and _scales. elif tensor_name.endswith("_scales"): blocks_name = tensor_name.replace("_scales", "_blocks") if blocks_name in weight_map: self._weight_scale_mapping[blocks_name] = tensor_name def _process_read_request( self, f: Any, req: ReadItem, planner: LoadPlanner ) -> None: """Override the Helper function that processes a single read request.""" tensor_fqn = req.storage_index.fqn # Check if this is a quantized tensor that needs dequantization if self._is_tensor_quantized(tensor_fqn): tensor = self._read_quantized_tensor_with_block_alignment(req, f) else: # Standard tensor reading slices = tuple( slice(offset, offset + length) for offset, length in zip(req.storage_offsets, req.lengths) ) tensor = f.get_slice(tensor_fqn)[slices] target_tensor = planner.resolve_tensor(req).detach() if target_tensor.size() != tensor.size(): raise AssertionError( f"req {req.storage_index} mismatch sizes {target_tensor.size()} vs {tensor.size()}" ) target_tensor.copy_(tensor) planner.commit_tensor(req, target_tensor) def _get_slice_to_block_mapping( self, req: ReadItem ) -> tuple[tuple[int, int], tuple[int, int], slice, slice]: """ Calculate which blocks correspond to the requested slice. Args: req: Read request containing tensor info and required slices Returns: Tuple of (row_block_range, col_block_range, row_slice, col_slice) """ # Get the slice information row_slice = slice( req.storage_offsets[0], req.storage_offsets[0] + req.lengths[0] ) col_slice = slice( req.storage_offsets[1], req.storage_offsets[1] + req.lengths[1] ) # Calculate which blocks this slice spans row_start_block = row_slice.start // self.block_size row_end_block = (row_slice.stop - 1) // self.block_size + 1 # Inclusive end col_start_block = col_slice.start // self.block_size col_end_block = (col_slice.stop - 1) // self.block_size + 1 # Inclusive end return ( (row_start_block, row_end_block), (col_start_block, col_end_block), row_slice, col_slice, ) def _dequantize_tensor_mxfp4( self, blocks: torch.Tensor, scales: torch.Tensor, req: ReadItem, group_start: int, offset_in_first_group: int, ) -> torch.Tensor: """ Dequantize a 4D tensor using MXFP4 format. Adapted from openai's implementation: https://github.com/openai/gpt-oss/blob/8890e95919f975a490fc0ba09ffb10890ec7319d/gpt_oss/torch/weights.py#L68 Args: blocks: Sliced quantized weight tensor of shape [a_slice, b_slice, groups_slice, B] in uint8 scales: FULL scale tensor of shape [a, b, c] in uint8 (will be converted to exponents) req: Read request containing slice information group_start: The starting group index in the checkpoint offset_in_first_group: Offset in values within the first group Returns: Dequantized tensor matching the requested shape """ # FP4 lookup table FP4_VALUES = [ +0.0, +0.5, +1.0, +1.5, +2.0, +3.0, +4.0, +6.0, -0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0, ] # blocks: [a_slice, b_slice, groups_slice, B] uint8. # Read slightly more groups than needed, and slice at the end. # Slice the scales to match the blocks dimensions. # [a_full, b_full, c_full] -> [a_slice, b_slice, groups_slice] dim0_start = req.storage_offsets[0] dim0_end = dim0_start + req.lengths[0] dim1_start = req.storage_offsets[1] dim1_end = dim1_start + req.lengths[1] num_groups = blocks.shape[2] scales = scales[ dim0_start:dim0_end, dim1_start:dim1_end, group_start : group_start + num_groups, ] scales = scales.to(torch.int32) - 127 assert blocks.shape[:-1] == scales.shape, ( f"{blocks.shape=} does not match {scales.shape=}" ) lut = torch.tensor(FP4_VALUES, dtype=self.target_dtype, device=blocks.device) *prefix_shape, G, B = blocks.shape rows_total = math.prod(prefix_shape) * G blocks = blocks.reshape(rows_total, B) scales = scales.reshape(rows_total, 1) out = torch.empty( rows_total, B * 2, dtype=self.target_dtype, device=blocks.device ) rows_per_chunk = 16384 * 512 for r0 in range(0, rows_total, rows_per_chunk): r1 = min(r0 + rows_per_chunk, rows_total) blk = blocks[r0:r1] exp = scales[r0:r1] # nibble indices -> int64 idx_lo = (blk & 0x0F).to(torch.long) idx_hi = (blk >> 4).to(torch.long) sub = out[r0:r1] sub[:, 0::2] = lut[idx_lo] sub[:, 1::2] = lut[idx_hi] torch.ldexp(sub, exp, out=sub) del idx_lo, idx_hi, blk, exp result = out.reshape(*prefix_shape, G, B * 2).view(*prefix_shape, G * B * 2) # Slice the last dimension to match the requested range. if offset_in_first_group > 0 or result.shape[-1] > req.lengths[2]: end_offset = offset_in_first_group + req.lengths[2] result = result[..., offset_in_first_group:end_offset] return result def _dequantize_tensor( self, weight: torch.Tensor, scale_inv: torch.Tensor, full_tensor_shape: torch.Size, slice_info: tuple[tuple[int, int], tuple[int, int], slice, slice], ) -> torch.Tensor: """ Dequantize a sliced tensor using the appropriate portion of the scale tensor. Args: weight: Sliced quantized weight tensor scale_inv: Full scale inverse tensor for dequantization full_tensor_shape: Shape of the original full tensor slice_info: Block mapping information from _get_slice_to_block_mapping Returns: Dequantized tensor """ (row_block_range, col_block_range, row_slice, col_slice) = slice_info # Convert to float32 for computation # Certain quantized dtypes like Float8_e4m3fn # don't support multiplication on CPU yet in PyTorch. upcasted_weight = weight.to(torch.float32) # Create output tensor in target dtype dequantized = weight.detach().to(dtype=self.target_dtype, copy=True) # Get the actual slice boundaries row_start_global = row_slice.start row_end_global = row_slice.stop col_start_global = col_slice.start col_end_global = col_slice.stop # Apply scaling factors to each block that intersects with our slice for block_i in range(row_block_range[0], row_block_range[1]): for block_j in range(col_block_range[0], col_block_range[1]): # Calculate the block boundaries in global coordinates block_row_start_global = block_i * self.block_size block_row_end_global = min( block_row_start_global + self.block_size, full_tensor_shape[0] ) block_col_start_global = block_j * self.block_size block_col_end_global = min( block_col_start_global + self.block_size, full_tensor_shape[1] ) # Find the intersection of the block with our slice intersect_row_start = max(block_row_start_global, row_start_global) intersect_row_end = min(block_row_end_global, row_end_global) intersect_col_start = max(block_col_start_global, col_start_global) intersect_col_end = min(block_col_end_global, col_end_global) # Skip if no intersection if ( intersect_row_start >= intersect_row_end or intersect_col_start >= intersect_col_end ): continue # Convert global coordinates to local coordinates in the sliced tensor local_row_start = intersect_row_start - row_start_global local_row_end = intersect_row_end - row_start_global local_col_start = intersect_col_start - col_start_global local_col_end = intersect_col_end - col_start_global # Get the block from the sliced tensor block = upcasted_weight[ local_row_start:local_row_end, local_col_start:local_col_end ] # Apply the scale factor scale = scale_inv[block_i, block_j] block = block * scale # Convert block to target dtype and store block_converted = block.to(dtype=self.target_dtype) dequantized[ local_row_start:local_row_end, local_col_start:local_col_end ] = block_converted return dequantized def _is_tensor_quantized(self, tensor_fqn: str) -> bool: """ Check if a tensor is a quantized. Args: tensor_fqn: Fully qualified name of the tensor Returns: True if tensor is quantized and has a corresponding scale tensor, False otherwise """ # Skip scale tensors themselves if tensor_fqn.endswith((".weight_scale_inv", "_scales")): return False # Check if this weight tensor has a corresponding scale tensor if tensor_fqn not in self._weight_scale_mapping: return False return True def _read_quantized_tensor_with_block_alignment( self, req: ReadItem, safetensor_file: Any ) -> torch.Tensor: """ Read a quantized tensor with block alignment. Args: req: Read request containing tensor info and required slices safetensor_file: Open safetensors file handle Returns: Dequantized tensor ready for use """ tensor_fqn = req.storage_index.fqn scale_fqn = self._weight_scale_mapping[tensor_fqn] try: group_start = 0 offset_in_first_group = 0 if tensor_fqn.endswith("_blocks"): # Full tensor is a 4D MXFP4 quantized tensor: [..., G, B]. # Each group G produces B * 2 dequantized values. # Checkpoint [..., G, B] -> dequantized [..., G*B*2]. # The planner gives 3D requests based on the dequantized shape. # Need to figure out which groups (dimension 2 in checkpoint) to read. # Use the quantized checkpoint shape to get the correct B. *prefix_shape, B = self._tensor_full_shapes[tensor_fqn + "_quantized"] values_per_group = B * 2 # Each byte has 2 nibbles (4-bit values). # Calculate which groups we need based on the requested range in dim 2. # Ensure the reequest is in 3D. assert len(req.storage_offsets) == 3 # Positions in dequantized space. dim2_start_deq = req.storage_offsets[2] dim2_length_deq = req.lengths[2] dim2_end_deq = dim2_start_deq + dim2_length_deq # Convert to group indices. group_start = dim2_start_deq // values_per_group group_end = (dim2_end_deq + values_per_group - 1) // values_per_group # Read only the necessary groups from checkpoint. weight_slices_4d = ( slice( req.storage_offsets[0], req.storage_offsets[0] + req.lengths[0] ), slice( req.storage_offsets[1], req.storage_offsets[1] + req.lengths[1] ), slice(group_start, group_end), slice(None), # Read all B values for each group. ) quantized_tensor = safetensor_file.get_slice(tensor_fqn)[ weight_slices_4d ] # Also track the offset within the first group offset_in_first_group = dim2_start_deq - ( group_start * values_per_group ) else: # 2D quantized tensor, use 2d block partition. weight_slices = tuple( slice(offset, offset + length) for offset, length in zip(req.storage_offsets, req.lengths) ) quantized_tensor = safetensor_file.get_slice(tensor_fqn)[weight_slices] # Load the corresponding scale inverse tensor (full tensor) scale_file_name = self._weight_map.get(scale_fqn) if scale_file_name is None: raise ValueError(f"Scale tensor {scale_fqn} not found in weight_map") # Check if scale tensor is in the same file as the weight tensor weight_file_name = self._weight_map.get(tensor_fqn) if scale_file_name == weight_file_name: # Scale tensor is in the same file, use current handle scale_inv = safetensor_file.get_tensor(scale_fqn) else: # Scale tensor is in a different file, need to open it from safetensors import safe_open # type: ignore[import] scale_file_path = Path(self.path) / scale_file_name with safe_open( scale_file_path, framework="pt", device="cpu" ) as scale_file: scale_inv = scale_file.get_tensor(scale_fqn) # Get the full tensor shape from our O(1) lookup cache full_tensor_shape = self._tensor_full_shapes.get(tensor_fqn) if full_tensor_shape is None: raise ValueError(f"Could not find full tensor shape for {tensor_fqn}") # Determine which dequantization function to use. if len(full_tensor_shape) == 2: # 2D block-wise quantization, e.g., used in deepseek v3.1 slice_info = self._get_slice_to_block_mapping(req) dequantized_tensor = self._dequantize_tensor( weight=quantized_tensor, scale_inv=scale_inv, full_tensor_shape=full_tensor_shape, slice_info=slice_info, ) elif tensor_fqn.endswith("_blocks"): # 4D with blocks along dimension 2, used in MXFP4, e.g. gpt-oss dequantized_tensor = self._dequantize_tensor_mxfp4( blocks=quantized_tensor, scales=scale_inv, req=req, group_start=group_start, offset_in_first_group=offset_in_first_group, ) else: raise ValueError("Unsupported quantization types") return dequantized_tensor except Exception as e: logger.error("Failed to read the quantized tensor!!") raise e