Add new cohere2_moe model#46115
Merged
Merged
Conversation
Port the Cohere2Moe (c4) model from cohere-transformers to upstream. Adds modular and generated modeling files, configuration, and unit tests. Registers Cohere2MoeConfig in CONFIG_MAPPING_NAMES, Cohere2MoeModel in MODEL_MAPPING_NAMES, and Cohere2MoeForCausalLM in MODEL_FOR_CAUSAL_LM_MAPPING. Adds "cohere2moe" -> "cohere2_moe" to SPECIAL_MODEL_TYPE_TO_MODULE_NAME so the auto-mapping resolves the module directory correctly. Co-authored-by: Cursor <[email protected]>
…e2Moe - Remove CohereTokenizer.__init__ and replace with convert_to_native_format classmethod so the tokenizer loads fully from tokenizer.json (v5-style). Adds add_bos_token/add_eos_token defaults to match 4.56.2.6 behaviour. - Update test_tokenization_cohere.py: disable test_tokenizer_from_extractor (no __init__ means extractor-based construction is unsupported), replace test_add_prefix_space_fast with tokenizer.json-based tests that verify pre_tokenizer/normalizer/decoder components are preserved on load. - Add Cohere2MoeVisionIntegrationTest to test_modeling_cohere2_vision.py with forward and generate tests for Command A+ (cohere2moe backbone). Co-authored-by: Cursor <[email protected]>
…e_position - Rename 'input_embeds' key to 'inputs_embeds' to match create_causal_mask() signature (typo introduced in generated modeling file). - Remove 'cache_position' from mask_kwargs; upstream create_causal_mask() does not accept this parameter (cohere-transformers-specific argument). Result: all 6 integration tests now pass (2 skipped: flash_attn not installed). Co-authored-by: Cursor <[email protected]>
Replace hardcoded expected outputs in test_model_flash_attn that were written for command_a+_bf16 with the correct outputs from mhlv2_bf16_clean. Co-authored-by: Cursor <[email protected]>
rope_scaling and sliding_window_pattern are consumed in __post_init__ and stored as derived attributes (rope_parameters and layer_types) which the modeling code reads directly. Co-authored-by: Cursor <[email protected]>
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Contributor
|
[For maintainers] Suggested jobs to run (before merge) run-slow: auto, cohere2, cohere2_moe, cohere2_vision |
Contributor
|
View the CircleCI Test Summary for this PR: |
5 tasks
yuchenxie4645
pushed a commit
to yuchenxie4645/transformers
that referenced
this pull request
May 28, 2026
* Add Cohere2Moe (cohere2_moe) model Port the Cohere2Moe (c4) model from cohere-transformers to upstream. Adds modular and generated modeling files, configuration, and unit tests. Registers Cohere2MoeConfig in CONFIG_MAPPING_NAMES, Cohere2MoeModel in MODEL_MAPPING_NAMES, and Cohere2MoeForCausalLM in MODEL_FOR_CAUSAL_LM_MAPPING. Adds "cohere2moe" -> "cohere2_moe" to SPECIAL_MODEL_TYPE_TO_MODULE_NAME so the auto-mapping resolves the module directory correctly. Co-authored-by: Cursor <[email protected]> * Drop CohereTokenizer.__init__; add vision integration tests for Cohere2Moe - Remove CohereTokenizer.__init__ and replace with convert_to_native_format classmethod so the tokenizer loads fully from tokenizer.json (v5-style). Adds add_bos_token/add_eos_token defaults to match 4.56.2.6 behaviour. - Update test_tokenization_cohere.py: disable test_tokenizer_from_extractor (no __init__ means extractor-based construction is unsupported), replace test_add_prefix_space_fast with tokenizer.json-based tests that verify pre_tokenizer/normalizer/decoder components are preserved on load. - Add Cohere2MoeVisionIntegrationTest to test_modeling_cohere2_vision.py with forward and generate tests for Command A+ (cohere2moe backbone). Co-authored-by: Cursor <[email protected]> * Fix create_causal_mask call: input_embeds -> inputs_embeds, drop cache_position - Rename 'input_embeds' key to 'inputs_embeds' to match create_causal_mask() signature (typo introduced in generated modeling file). - Remove 'cache_position' from mask_kwargs; upstream create_causal_mask() does not accept this parameter (cohere-transformers-specific argument). Result: all 6 integration tests now pass (2 skipped: flash_attn not installed). Co-authored-by: Cursor <[email protected]> * Update flash_attn expected texts for mhlv2_bf16_clean model Replace hardcoded expected outputs in test_model_flash_attn that were written for command_a+_bf16 with the correct outputs from mhlv2_bf16_clean. Co-authored-by: Cursor <[email protected]> * update path * address comments * fix deprecate * address comments * address comments * update mapping * fix ci * fix ci * format * fix ci * format * Add Cohere2MoeConfig to check_config_attributes ignore list rope_scaling and sliding_window_pattern are consumed in __post_init__ and stored as derived attributes (rope_parameters and layer_types) which the modeling code reads directly. Co-authored-by: Cursor <[email protected]> * add date? * Update docs/source/en/model_doc/cohere2_moe.md * a few improvements * extract the router explicitly * simplify attention * attributes * fix the output type * simplify tests * fix init weights * revert changes to tokenizer - will be in hub config * style * mapping already exist, use it directly --------- Co-authored-by: Terrencezzj <[email protected]> Co-authored-by: Cursor <[email protected]> Co-authored-by: bharat <[email protected]>
kashif
pushed a commit
to kashif/transformers
that referenced
this pull request
Jun 1, 2026
* Add Cohere2Moe (cohere2_moe) model Port the Cohere2Moe (c4) model from cohere-transformers to upstream. Adds modular and generated modeling files, configuration, and unit tests. Registers Cohere2MoeConfig in CONFIG_MAPPING_NAMES, Cohere2MoeModel in MODEL_MAPPING_NAMES, and Cohere2MoeForCausalLM in MODEL_FOR_CAUSAL_LM_MAPPING. Adds "cohere2moe" -> "cohere2_moe" to SPECIAL_MODEL_TYPE_TO_MODULE_NAME so the auto-mapping resolves the module directory correctly. Co-authored-by: Cursor <[email protected]> * Drop CohereTokenizer.__init__; add vision integration tests for Cohere2Moe - Remove CohereTokenizer.__init__ and replace with convert_to_native_format classmethod so the tokenizer loads fully from tokenizer.json (v5-style). Adds add_bos_token/add_eos_token defaults to match 4.56.2.6 behaviour. - Update test_tokenization_cohere.py: disable test_tokenizer_from_extractor (no __init__ means extractor-based construction is unsupported), replace test_add_prefix_space_fast with tokenizer.json-based tests that verify pre_tokenizer/normalizer/decoder components are preserved on load. - Add Cohere2MoeVisionIntegrationTest to test_modeling_cohere2_vision.py with forward and generate tests for Command A+ (cohere2moe backbone). Co-authored-by: Cursor <[email protected]> * Fix create_causal_mask call: input_embeds -> inputs_embeds, drop cache_position - Rename 'input_embeds' key to 'inputs_embeds' to match create_causal_mask() signature (typo introduced in generated modeling file). - Remove 'cache_position' from mask_kwargs; upstream create_causal_mask() does not accept this parameter (cohere-transformers-specific argument). Result: all 6 integration tests now pass (2 skipped: flash_attn not installed). Co-authored-by: Cursor <[email protected]> * Update flash_attn expected texts for mhlv2_bf16_clean model Replace hardcoded expected outputs in test_model_flash_attn that were written for command_a+_bf16 with the correct outputs from mhlv2_bf16_clean. Co-authored-by: Cursor <[email protected]> * update path * address comments * fix deprecate * address comments * address comments * update mapping * fix ci * fix ci * format * fix ci * format * Add Cohere2MoeConfig to check_config_attributes ignore list rope_scaling and sliding_window_pattern are consumed in __post_init__ and stored as derived attributes (rope_parameters and layer_types) which the modeling code reads directly. Co-authored-by: Cursor <[email protected]> * add date? * Update docs/source/en/model_doc/cohere2_moe.md * a few improvements * extract the router explicitly * simplify attention * attributes * fix the output type * simplify tests * fix init weights * revert changes to tokenizer - will be in hub config * style * mapping already exist, use it directly --------- Co-authored-by: Terrencezzj <[email protected]> Co-authored-by: Cursor <[email protected]> Co-authored-by: bharat <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
As per the title!