[convert] Support for DeepSeek-V3.2 and Dequantizing

[brian-dellabetta]

Corequisite:

• DeepSeek V3.2 support llm-compressor#2491

This PR enhances the convert_checkpoint entrypoint to handle dequantization as well, and adds new functionality to be compatible with DeepSeek-V3.2:

• Add a FP8Dequantizer` converter to upconvert checkpoints in quant method fp8 and quant scheme FP8_BLOCK back to bfloat16
• Add get_dependency_weight method to Converter interface, so converters can define if a weight has dependency weights that also need to be processed along with.
• Add build_inverse_weight_maps logic, following pattern in llm-compressor, so that all weight dependencies can be loaded in with the given weight, even if they live in separate safetensors files. The deepseek model often splits a module's weight and weight_scale_inv tensors across different files, and they need to be processed together when dequantizing.
• Update converter create_quant_config signature to return an Optional field. This is needed when converting from one checkpoint to bfloat16, when no accompanying compressed-tensors quantization config is needed.

Merge in conjunction with vllm-project/llm-compressor#2491

TODOs:

• rebase against [MTP] Add functionality to save mtp layers from the original model + refactor safetensors utilities #640
• refactor to align with [Distributed] [model_free_ptq] Eliminate reindexing step via fine-grained parallelized partial reads llm-compressor#2498

Test Plan:

• unit tests for new converter and new reindex entrypoint
• nightly tests for new converter on accompanying llm-compressor PR

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to adddress the lint failures. You will need to install the
dev optional install to get the required linting packages.

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to adddress the lint failures. You will need to install the
dev optional install to get the required linting packages.

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to adddress the lint failures. You will need to install the
dev optional install to get the required linting packages.

[kylesayrs]

Outside of error for single-shard models, looks good to me! Thanks for being open to suggestions

[kylesayrs]

Suggested change

	all_dependencies: set[str] = set()
	for values in weight_deps_dict.values():
	for value in values:
	all_dependencies.add(value)
	all_dependencies: set[str] = set().union(*weight_deps_dict.values())

[kylesayrs]

Please give an example of an "optional dependency" to make the concept clear

[kylesayrs]

Suggested change

	weight_deps_dict: dict[str, set[str]] = defaultdict(set)
	weight_deps_dict: dict[str, dict[str, bool]] = dict()

[kylesayrs]

Suggested change

	resolved_path = model_files.get(weight_to_add_shard_name)
	resolved_path = model_files[weight_to_add_shard_name]

[kylesayrs]

ty

[kylesayrs]

Shouldn't this be an error?

[kylesayrs]

This makes sense for now, but is inflexible if we want to support mixed recipes. For example, convert some weights to full precision, but convert other weights from fp8 to compressed-tensors

[kylesayrs]

This will error if there is no index file. Use something like this instead

def get_weight_map(model_files):
  index_file = find_safetensors_index_file(model_files)
  if index_file is not None:
    with open(index_file, "r") as f:
        return json.load(f)["weight_map"]
  else:
    with safe_open(SAFE_WEIGHTS_NAME, "r") as file:
      return {tensor: SAFE_WEIGHTS_NAME for tensor in file.keys()}

[kylesayrs]

qconfig field is not guaranteed to exist

Suggested change

	if quant_config_data is None:
	if quant_config_data is None and QUANTIZATION_CONFIG_NAME in config_data:

[kylesayrs]

Suggested change

	if file_path.endswith("safetensors.index.json"):
	if file_path.endswith(SAFE_WEIGHTS_INDEX_NAME):