Class: Cohere::Transcribe::Runtime::ModelProvider

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

Overview

Resolves native Transformers Dense checkpoints, downloads only their required artifacts, converts them once to CrispASR GGUF, and retains no Python-side state.

Constant Summary collapse

CACHE_LAYOUT_VERSION =

Increment whenever conversion semantics or the native tensor types written to GGUF change. Version 2 adds first-class BF16 artifacts.

2

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(hub: Hub.new, cache_dir: nil, native_session_class: nil, converter: nil) ⇒ ModelProvider

Returns a new instance of ModelProvider.



21
22
23
24
25
26
27
28
29
 
# File 'lib/cohere/transcribe/runtime/model_provider.rb', line 21

def initialize(hub: Hub.new, cache_dir: nil, native_session_class: nil,
               converter: nil)
  @hub = hub
  root = cache_dir || ENV["COHERE_TRANSCRIBE_CACHE"] ||
         File.join(ENV.fetch("XDG_CACHE_HOME", File.expand_path("~/.cache")), "cohere-transcribe")
  @cache_dir = Pathname(root).expand_path
  @native_session_class = native_session_class
  @converter = converter
end

Instance Attribute Details

#cache_dir ⇒ Object (readonly)

Returns the value of attribute cache_dir.



19
20
21
 
# File 'lib/cohere/transcribe/runtime/model_provider.rb', line 19

def cache_dir
  @cache_dir
end

#hub ⇒ Object (readonly)

Returns the value of attribute hub.



19
20
21
 
# File 'lib/cohere/transcribe/runtime/model_provider.rb', line 19

def hub
  @hub
end

Instance Method Details

#converted_model_path(identity, options) ⇒ Object 



63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
 
# File 'lib/cohere/transcribe/runtime/model_provider.rb', line 63

def converted_model_path(identity, options)
  model_directory = materialize_source(identity)
  output_type = { "fp32" => :f32, "bf16" => :bf16 }.fetch(options.dtype, :f16)
  fingerprint = source_fingerprint(identity, model_directory, output_type)
  directory = conversion_cache_directory
  output = directory.join("#{fingerprint}-#{output_type}.gguf")
  return output if valid_gguf?(output, fingerprint: fingerprint, output_type: output_type)

  lock_path = Pathname("#{output}.lock")
  open_conversion_lock(lock_path) do |lock|
    raise TranscriptionRuntimeError, "Cannot acquire Dense conversion lock for #{output}" unless lock.flock(File::LOCK_EX)
    return output if valid_gguf?(output, fingerprint: fingerprint, output_type: output_type)

    cleanup_conversion_temporaries(output)
    remove_cache_entry(output)
    remove_cache_entry(conversion_marker(output))
    begin
      converter.convert(
        model_dir: model_directory,
        output_path: output,
        output_type: output_type,
        overwrite: false,
        fsync: true
      )
      unless source_fingerprint(identity, model_directory, output_type) == fingerprint
        raise TranscriptionRuntimeError,
              "Dense checkpoint changed during Dense conversion; retry with a stable model directory"
      end
      write_conversion_marker(output, fingerprint: fingerprint, output_type: output_type)
    rescue Exception # rubocop:disable Lint/RescueException -- clean partial artifacts before propagating interrupts
      remove_cache_entry(output)
      remove_cache_entry(conversion_marker(output))
      raise
    end
  end
  unless valid_gguf?(output, fingerprint: fingerprint, output_type: output_type)
    raise TranscriptionRuntimeError,
          "Dense conversion did not produce a valid GGUF file at #{output}"
  end

  output
rescue DenseConverter::Error, Safetensors::Error, PyTorchCheckpoint::Error, GGUF::Error => e
  raise TranscriptionRuntimeError, "Cannot convert Dense Cohere checkpoint: #{e.message}"
rescue Hub::Error => e
  raise TranscriptionRuntimeError, e.message
rescue SystemCallError => e
  raise TranscriptionRuntimeError, "Cannot prepare native Dense model: #{e.message}"
end

#materialize_source(identity) ⇒ Object

Raises:



112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
 
# File 'lib/cohere/transcribe/runtime/model_provider.rb', line 112

def materialize_source(identity)
  local = Pathname(identity.model_id).expand_path
  return local.realpath if local.directory?

  revision = identity.model_revision
  raise TranscriptionRuntimeError, "Remote model identity has no immutable revision" unless revision

  DenseConverter::REQUIRED_ARTIFACT_FILENAMES.each do |filename|
    hub.download(identity.model_id, filename, revision: revision)
  end
  snapshot = hub.snapshot_path(identity.model_id, revision)
  return snapshot if hub.cached_file(identity.model_id, "model.safetensors", revision: revision)
  return snapshot if hub.cached_file(identity.model_id, "pytorch_model.bin", revision: revision)

  safetensors_index = hub.cached_file(
    identity.model_id, "model.safetensors.index.json", revision: revision
  )
  if safetensors_index
    download_index_shards(identity, safetensors_index, extension: ".safetensors", label: "Safetensors")
    return snapshot
  end
  pytorch_index = hub.cached_file(
    identity.model_id, "pytorch_model.bin.index.json", revision: revision
  )
  if pytorch_index
    download_index_shards(identity, pytorch_index, extension: ".bin", label: "PyTorch")
    return snapshot
  end

  files = hub.list_files(identity.model_id, revision: revision)
  if files.include?("model.safetensors")
    hub.download(identity.model_id, "model.safetensors", revision: revision)
  elsif files.include?("model.safetensors.index.json")
    index = hub.download(identity.model_id, "model.safetensors.index.json", revision: revision)
    download_index_shards(identity, index, extension: ".safetensors", label: "Safetensors")
  elsif files.include?("pytorch_model.bin")
    hub.download(identity.model_id, "pytorch_model.bin", revision: revision)
  elsif files.include?("pytorch_model.bin.index.json")
    index = hub.download(identity.model_id, "pytorch_model.bin.index.json", revision: revision)
    download_index_shards(identity, index, extension: ".bin", label: "PyTorch")
  else
    raise TranscriptionRuntimeError,
          "#{identity.model_id}@#{revision} has no supported Dense Safetensors or PyTorch weights"
  end
  snapshot
end

#open(identity, options) ⇒ Object 



55
56
57
58
59
60
61
 
# File 'lib/cohere/transcribe/runtime/model_provider.rb', line 55

def open(identity, options)
  ModelIdentity.verify_model_weight_artifacts(identity, hub: hub)
  path = converted_model_path(identity, options)
  (@native_session_class || ASR::NativeSession).new(path, options)
rescue ArgumentError, TypeError, Hub::Error => e
  raise TranscriptionRuntimeError, e.message
end

#resolve(options, verify_model_weights: false) ⇒ Object

Planning may discover that every requested output can be resumed from a checkpoint. Defer multi-gigabyte weight discovery until #open unless a caller explicitly asks to validate inference artifacts now.



34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
 
# File 'lib/cohere/transcribe/runtime/model_provider.rb', line 34

def resolve(options, verify_model_weights: false)
  identity = ModelIdentity.resolve(
    path_text(options.model),
    options.model_revision,
    path_text(options.adapter),
    options.adapter_revision,
    hub: hub,
    verify_weight_artifacts: verify_model_weights
  )
  unless identity.model_format == :dense
    raise TranscriptionConfigurationError,
          "Saved #{identity.model_format} checkpoints are outside the core Dense Ruby inference path"
  end
  if identity.adapter_id
    raise TranscriptionConfigurationError,
          "PEFT/LoRA adapters are not supported by the native Ruby Dense runtime"
  end

  identity
end