Class: Cohere::Transcribe::PyTorchCheckpoint::Reader

Inherits:
Object
  • Object
show all
Defined in:
lib/cohere/transcribe/pytorch_checkpoint.rb

Overview

One checkpoint shard. Tensor bytes are never materialized as a whole.

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(path) ⇒ Reader

Returns a new instance of Reader.



782
783
784
785
786
787
788
789
# File 'lib/cohere/transcribe/pytorch_checkpoint.rb', line 782

def initialize(path)
  @path = Pathname(path).expand_path
  @storages = {}
  @storage_payloads = {}
  @verified_storage_payloads = {}
  @storage_verification_mutex = Mutex.new
  @tensors = read_checkpoint.freeze
end

Instance Attribute Details

#pathObject (readonly)

Returns the value of attribute path.



780
781
782
# File 'lib/cohere/transcribe/pytorch_checkpoint.rb', line 780

def path
  @path
end

#tensorsObject (readonly)

Returns the value of attribute tensors.



780
781
782
# File 'lib/cohere/transcribe/pytorch_checkpoint.rb', line 780

def tensors
  @tensors
end

Instance Method Details

#fetch(name) ⇒ Object



799
800
801
802
803
# File 'lib/cohere/transcribe/pytorch_checkpoint.rb', line 799

def fetch(name)
  tensors.fetch(name)
rescue KeyError
  raise Error, "Tensor #{name.inspect} is not present in #{path}"
end

#key?(name) ⇒ Boolean

Returns:

  • (Boolean)


795
796
797
# File 'lib/cohere/transcribe/pytorch_checkpoint.rb', line 795

def key?(name)
  tensors.key?(name)
end

#namesObject



791
792
793
# File 'lib/cohere/transcribe/pytorch_checkpoint.rb', line 791

def names
  tensors.keys
end

#write_tensor(tensor, output, target_dtype:, converter: Safetensors::DTypeConverter.default, chunk_bytes: Safetensors::DEFAULT_CHUNK_BYTES) ⇒ Object

Raises:

  • (ArgumentError)


805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
# File 'lib/cohere/transcribe/pytorch_checkpoint.rb', line 805

def write_tensor(tensor, output, target_dtype:, converter: Safetensors::DTypeConverter.default,
                 chunk_bytes: Safetensors::DEFAULT_CHUNK_BYTES)
  raise ArgumentError, "Tensor #{tensor.name.inspect} belongs to a different PyTorch reader" unless tensor.reader.equal?(self)

  target_dtype = target_dtype.to_s.upcase
  unless FLOAT_DTYPES.include?(tensor.dtype) && FLOAT_DTYPES.include?(target_dtype)
    raise Error, "Cannot convert #{tensor.name.inspect} from #{tensor.dtype} to #{target_dtype}"
  end
  raise ArgumentError, "chunk_bytes must be positive" unless chunk_bytes.positive?

  verify_storage_payload!(tensor)

  source_width = Safetensors::DTYPE_BYTES.fetch(tensor.dtype)
  elements_per_chunk = [chunk_bytes / source_width, 1].max
  bytes_per_chunk = elements_per_chunk * source_width
  unless contiguous?(tensor)
    return write_strided_tensor(
      tensor, output, target_dtype: target_dtype, converter: converter,
                      elements_per_chunk: elements_per_chunk, source_width: source_width
    )
  end

  remaining = tensor.nbytes
  written = 0
  File.open(path, "rb") do |source|
    source.seek(tensor.data_start)
    while remaining.positive?
      requested = [remaining, bytes_per_chunk].min
      chunk = source.read(requested)
      raise Error, "Unexpected end of #{path} while reading #{tensor.name.inspect}" if chunk.nil? || chunk.bytesize != requested

      converted = converter.convert(chunk, from: tensor.dtype, to: target_dtype)
      output.write(converted)
      written += converted.bytesize
      remaining -= requested
    end
  end
  expected = tensor.element_count * Safetensors::DTYPE_BYTES.fetch(target_dtype)
  return written if written == expected

  raise Error, "Converted tensor #{tensor.name.inspect} wrote #{written} bytes; expected #{expected}"
end