Add native DeepSeek-V3.2 support

[XingyuHu109]

Summary

This PR adds native Transformers support for DeepSeek-V3.2.

It introduces a new deepseek_v32 model family so the official checkpoints resolve through the standard auto classes without trust_remote_code. The implementation keeps the DeepSeek-V3 MoE structure and plugs in the in-tree DSA attention/indexer path that V3.2 uses. Docs and a dedicated test module are included as well.

Closes #41196.

Validation

• PYTHONPATH=src python -m pytest tests/models/deepseek_v32/test_modeling_deepseek_v32.py -q
  • 123 passed, 129 skipped
• PYTHONPATH=src python -m ruff check src/transformers/models/deepseek_v32 tests/models/deepseek_v32
  • passed
• official auto-class resolution with this branch:
  • AutoConfig.from_pretrained("deepseek-ai/DeepSeek-V3.2") -> DeepseekV32Config
  • AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V3.2", ...) -> DeepseekV32ForCausalLM
• end-to-end native load and generation against the published deepseek-ai/DeepSeek-V3.2 checkpoint completed successfully
• current GitHub checks are green

Note

The public tokenizer config does not currently ship a chat_template, so apply_chat_template(...) still needs an explicit template.

[Copilot]

Pull request overview

This PR adds native Hugging Face Transformers support for the new DeepSeek-V3.2 architecture (deepseek_v32) so official checkpoints (e.g. deepseek-ai/DeepSeek-V3.2) resolve through standard auto-classes without requiring trust_remote_code.

Changes:

• Introduces the deepseek_v32 model family (config + PyTorch modeling), registers it in auto-class mappings, and adds docs/tests.
• Improves robustness of dynamic attention/experts-implementation detection (avoids KeyError when sys.modules lacks the model module entry).
• Extends accelerate integration to support preload_module_classes dispatch (relevant for disk-offloaded MoE blocks) and fixes FP8 quantizer device-map validation precedence.

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/transformers/models/deepseek_v32/modular_deepseek_v32.py	Modular source defining DeepSeek-V3.2 config/model with DSA indexer + MoE.
src/transformers/models/deepseek_v32/configuration_deepseek_v32.py	Generated config for deepseek_v32.
src/transformers/models/deepseek_v32/modeling_deepseek_v32.py	Generated PyTorch implementation + generation support.
src/transformers/models/deepseek_v32/__init__.py	Lazy import plumbing for new model package.
src/transformers/models/auto/configuration_auto.py	Registers deepseek_v32 config in auto config mapping + display name.
src/transformers/models/auto/modeling_auto.py	Registers deepseek_v32 model + causal LM in auto model mappings.
src/transformers/models/__init__.py	Exposes the new model package under transformers.models.
tests/models/deepseek_v32/test_modeling_deepseek_v32.py	New unit tests validating config fields, auto-class resolution, caching shapes, and disk-offloaded MoE behavior.
docs/source/en/model_doc/deepseek_v32.md	New model documentation page + usage example.
docs/source/en/_toctree.yml	Adds DeepSeek-V3.2 docs page to the docs navigation.
src/transformers/integrations/accelerate.py	Adds preload_module_classes dispatch support with a compatibility shim for hook recursion.
src/transformers/modeling_utils.py	Makes _can_set_attn_implementation / _can_set_experts_implementation robust to missing sys.modules entries.
src/transformers/quantizers/quantizer_finegrained_fp8.py	Fixes device-map validation precedence and updates isinstance check style.
tests/quantization/finegrained_fp8/test_fp8.py	Adds test ensuring pre-quantized FP8 models allow disk offload in device maps.
tests/utils/test_modeling_utils.py	Adds regression test for missing sys.modules module-cache entry.
utils/check_config_attributes.py	Allows V3.2 “metadata” fields not referenced by modeling code.
src/transformers/conversion_mapping.py / docs/source/en/weightconverter.md	Adds deepseek_v32 to weight conversion pattern mapping.

[Copilot]

In _attach_align_device_hook_on_blocks_for_preload, the early-return branch checks not isinstance(offload, dict), but later the function treats offload as a generic Mapping. If accelerate ever passes a non-dict mapping (e.g. OrderedDict) this condition will be true and the code will incorrectly treat the mapping like a scalar boolean, potentially attaching the wrong hooks. Use Mapping consistently here (i.e. not isinstance(offload, Mapping)) so the scalar-vs-mapping logic is correct.

Suggested change

	if not isinstance(execution_device, Mapping) and not isinstance(offload, dict):
	if not isinstance(execution_device, Mapping) and not isinstance(offload, Mapping):

[Copilot]

The docstring example references a model repo (meta-deepseek_v32/DeepseekV32-2-7b-hf) that does not match the official checkpoint this PR adds native support for (deepseek-ai/DeepSeek-V3.2). This will mislead users copying the snippet; please update the example to use the official model id (and matching tokenizer).

Suggested change

	>>> model = DeepseekV32ForCausalLM.from_pretrained("meta-deepseek_v32/DeepseekV32-2-7b-hf")
	>>> tokenizer = AutoTokenizer.from_pretrained("meta-deepseek_v32/DeepseekV32-2-7b-hf")
	>>> model = DeepseekV32ForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V3.2")
	>>> tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3.2")

[Copilot]

Several tests are skipped with reasons phrased as uncertainty (e.g. "Not sure MoE can pass this..."). Skip reasons should be specific and actionable (ideally referencing a known limitation or a tracking issue) so it’s clear whether this is an expected permanent limitation or a temporary gap to fix.

Suggested change

	@unittest.skip("Not sure MoE can pass this + indexer outputs are not deterministic wrt padding")
	def test_left_padding_compatibility(self):
	pass
	@unittest.skip("Not sure MoE can pass this + indexer outputs are not deterministic wrt padding")
	def test_sdpa_padding_matches_padding_free_with_position_ids(self):
	pass
	@unittest.skip("Not sure MoE can pass this + indexer outputs are not deterministic wrt padding")
	@unittest.skip("Skipped: MoE routing with the DSA indexer produces non-deterministic outputs with respect to padding, so left-padding compatibility cannot be reliably tested.")
	def test_left_padding_compatibility(self):
	pass
	@unittest.skip("Skipped: MoE routing with the DSA indexer produces non-deterministic outputs with respect to padding, so SDPA vs padding-free behavior with position_ids cannot be reliably compared.")
	def test_sdpa_padding_matches_padding_free_with_position_ids(self):
	pass
	@unittest.skip("Skipped: MoE routing with the DSA indexer produces non-deterministic outputs with respect to padding, making this overfitting test unreliable.")

[XingyuHu109]

All checks are green now, and the DeepSeek-V3.2 native load/generation path has been validated on this branch. Would appreciate a review when someone working in this area has time.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto, deepseek_v32, finegrained_fp8

[Rocketknight1]

cc @ArthurZucker!