* upgrade to torchao 0.17.0
* chore: lint
* refactor attention handling
* replace legacy attention boolean flags with capability properties
Replace checks with capability-based properties derived from attn_implementation
This separates three concerns that were conflated under flash_attention:
1. Backend selection -> attn_implementation enum
2. Packing capability -> attn_supports_packing property
3. Flash-attn library dependency -> attn_uses_flash_lib property
* compute attn capability flags in normalizer instead of properties
* make attn_implementation the single source of truth
* move attention-dependent validators to mode=after
* migrate remaining consumers to canonical attn_implementation
* expand attention tests + rewrite docs
* migrate example configs to canonical attn_implementation
* update doc snippets + reject gemma4-hybrid with non-FA2 backend
* remove dead gemma4 branch in _set_attention_config
* fix duplicate attn_implementation in gpt-oss yamls and flaky caplog tests
* drop "Phase 2" naming from attn-implementation tests
* regroup attn_implementation tests by feature concern
* clean up verbose comments and remove MD
Signed-off-by: Wing Lian <wing@axolotl.ai>
Co-authored-by: Axolotl Swarm <no-reply@axolotl.ai>
* fix(collator): pass return_dict=True at apply_chat_template top level for transformers 5.x
In transformers 5.x, ProcessorMixin.apply_chat_template gained its own
`return_dict` parameter (defaulting to False). When return_dict=False
and tokenize=True the method returns out["input_ids"] directly — a 2-D
tensor — rather than the full BatchFeature dict.
The old code placed `return_dict=True` inside processor_kwargs. In
transformers 5.x those kwargs are forwarded to the underlying processor
call self(...) where _merge_kwargs silently ignores any key not present
in MllamaProcessorKwargs (emitting a warning). The outer return_dict
therefore stayed False, apply_chat_template returned the raw input_ids
tensor, and the subsequent `batch["input_ids"]` attempted to index a
2-D tensor with the 9-character string "input_ids", producing:
IndexError: too many indices for tensor of dimension 2
The fix is to pass return_dict=True as a top-level keyword argument to
apply_chat_template (where it is actually consumed) and remove it from
processor_kwargs (where it was silently dropped). No version guard is
needed: transformers is pinned to ==5.5.4 in pyproject.toml.
Adds a unit-level regression test (tests/test_mm_chat_collator.py) that
mocks the processor to return a raw tensor when apply_chat_template is
called without top-level return_dict=True, verifying the four invariants:
process_rows returns a dict, input_ids is 2-D, labels is 2-D, and
apply_chat_template receives return_dict=True as a top-level kwarg.
Fixes: tests/e2e/test_llama_vision.py::TestLlamaVision::test_lora_llama_vision_multimodal_dataset
Fixes: tests/e2e/test_llama_vision.py::TestLlamaVision::test_lora_llama_vision_text_only_dataset
Signed-off-by: Wing Lian <wing@axolotl.ai>
Co-authored-by: Axolotl Swarm <no-reply@axolotl.ai>
* fix(collator): process_rows returns dict (BatchFeature) shape
Two related changes for the multimodal chat collator under transformers 5.x:
1. Wrap apply_chat_template result in dict(...) so process_rows returns
a plain dict rather than a BatchFeature instance. BatchFeature is a
Mapping but not a dict; downstream code that did
batch["labels"] = self.processing_strategy.process_labels(batch["input_ids"])
would index on a tensor when the result wasn't dict-shaped, raising
IndexError: too many indices for tensor of dimension 2
2. Soften the regression test's contract from `dict` to `Mapping` so it
exercises the actual semantic guarantee (key/value access) rather
than the implementation detail (dict vs BatchFeature). Test guards
against the original transformers 5.x breakage where apply_chat_template's
return_dict default went from True to False.
Includes regression test under tests/test_mm_chat_collator.py.
Bug surfaced via swarm dispatch task_01KQHPNAYD8XARSNSDJVW1GPF6 against
attn-implementation-refactor; squash-merged from agent commits 4de886fd
+ dc9fcf4f.
Signed-off-by: Wing Lian <wing@axolotl.ai>
---------
Signed-off-by: Wing Lian <wing@axolotl.ai>
Co-authored-by: Axolotl Swarm <no-reply@axolotl.ai>
* feat: LoRA kernel support for bias, dropout, dora, embeddings
* chore: lint
* chore: lint
* address PR feedback, add regression tests, add fsdp2 tests for lora kernels
* update tests for new sigs
* update tests now that bias and dropout are supported
* upgrade transformers to 5.1.0 and torchao to 0.16.0
* upgrade trl for parity
* handle trl api changes
* orpo doesn't have max_prompt_len to check anymore
* cpoconfig doesn't take max_prompt_length and fix cpu offload
* slow fsdp1 test
* triton min 3.4.0 and liger to 0.7.0
* use transformers main for now for zero3 fix
* handle group_by_length change
* fix changes upstream
* mark skip flaky test
* use transformers latest release 5.2.0
* Prepare for transformers v5 upgrade
* fix hf cli
* update for hf hub changes
* fix tokenizer apply_chat_template args
* remap include_tokens_per_second
* fix tps
* handle migration for warmup
* use latest hf hub
* Fix scan -> ls
* fix import
* fix for renaming of mistral common tokenizer -> backend
* update for fixed tokenziation for llama
* Skip phi35 tests for now
* remove mistral patch fixed upstream in huggingface/transformers#41439
* use namespacing for patch
* don't rely on sdist for e2e tests for now
* run modal ci without waiting too
* Fix dep for ci
* fix imports
* Fix fp8 check
* fsdp2 fixes
* fix version handling
* update fsdp version tests for new v5 behavior
* Fail multigpu tests after 3 failures
* skip known v5 broken tests for now and cleanup
* bump deps
* unmark skipped test
* re-enable test_fsdp_qlora_prequant_packed test
* increase multigpu ci timeout
* skip broken gemma3 test
* reduce timout back to original 120min now that the hanging test is skipped
* fix for un-necessary collator for pretraining with bsz=1
* fix: safe_serialization deprecated in transformers v5 rc01 (#3318)
* torch_dtype deprecated
* load model in float32 for consistency with tests
* revert some test fixtures back
* use hf cache ls instead of scan
* don't strip fsdp_version
more fdsp_Version fixes for v5
fix version in fsdp_config
fix aliasing
fix fsdp_version check
check fsdp_version is 2 in both places
* Transformers v5 rc2 (#3347)
* bump dep
* use latest fbgemm, grab model config as part of fixture, un-skip test
* import AutoConfig
* don't need more problematic autoconfig when specifying config.json manually
* add fixtures for argilla ultrafeedback datasets
* download phi4-reasoning
* fix arg
* update tests for phi fast tokenizer changes
* use explicit model types for gemma3
---------
Co-authored-by: Wing Lian <wing@axolotl.ai>
* fix: AutoModelForVision2Seq -> AutoModelForImageTextToText
* chore: remove duplicate
* fix: attempt fix gemma3 text mode
* chore: lint
* ga release of v5
* need property setter for name_or_path for mistral tokenizer
* vllm not compatible with transformers v5
* setter for chat_template w mistral too
---------
Co-authored-by: NanoCode012 <nano@axolotl.ai>
Co-authored-by: salman <salman.mohammadi@outlook.com>
* make sure to use ray prepare for dataloader fixes
* ray tests use 2.7.0+
* don't call init_distributed w ray and deepspeed
* handle dict deepspeed config
* better handling of dict deepspeed config
* use json.dumps
* guard to_dict
* wrap import for optional ray
* upgrade transformers to 4.57.0
* remove deprecated autoawq and use latest peft
* remove autoawq from setuptools script
* fix imports
* make sure torchvision is installed
* remove support for BetterTransformer
* skip fsdp_qlora_prequant test
* more robust error reporting
* checkpoint model on first step callback
* remove debug
* add test cases; update existing tests not to save on first step
* move test out of solo
* delete
* default to False
* typo
* upgrade peft to 0.16.0
* upgrade datasets to 4.0.0
* refactor dupes from merge/rebase
* fix check for fsdp1 + sharded_state_dict
* use full state dict for ci
* upgrade trl==0.19.1
* add vllm for tests for grpo
* fixes to work with latest trl
* need data_parallel_size config too
* support for vllm_mode for server / colocate
* vllm settings for colocate
* relax vllm version
* bump min hf hub for latest vllm support
* add hints on string literal for vllm mode
* use latest transformers 4.53.2
* tweak acceptable loss on flaky test_ds_zero3_packed test
* don't run flaky vllm/grpo tests for now
* FSDP2 args migration implementation
This commit implements the migration to FSDP2 arguments including:
- FSDP2 support with LoRA training
- DPO integration with FSDP2
- Model loading fixes and refactoring
- CPU offloading and PEFT handling
- Test updates and CI improvements
- Bug fixes for dtype errors and various edge cases
* update transformers to 4.53.0
* remove attention_mask from signature columns if using packing
* remove attention_mask column from dataloader
* update signature of flash attn forward for ring attn patch
* fix FSDP
* patch ring-flash-attn with upstream signature fix
* fix patch indentation level
* fix the patch
* add batch flattening smoke test with loss check that works in older transformers
* fix patch
* don't drop attention mask for flex
* more fixes
* patch create_causal_mask for packing w flex
* global torch manual_seed fixture
* tweak loss checks
* fix patch and use single batch for flex
* don't need to reload
* fix causal mask patch
* use transformers patch releasE
* make sure env var is string
* make sure to drop attention mask for flex w packing for latest transformers patch release
* tweak loss
* guard on signature columns before removing attention mask
* bump loss
* set remove isn't chainable
* skip slow mistral test in 2.5.1
* bump hf deps
* upgrade liger-kernel too
* install cce from fork for transformers fix
* fix reference to vocab size in gemma3 patch
* use padding_idx instead of pad_token_id
* remove fixed gemma3 patch
* use updated cce fork
* fix local mllama cce patches w docstring
* add test for multipack with trainer setup and fix trainer for trainer refactor upstream
* bump modal version
* guard for iterable datasetS
* mllama model arch layout changed in latest transformers
* fix batch sampler with drop_last
* fix: address upstream vlm changes for lora
* fix: update references to old lora target path
* fix: remove mllama fa2 patch due to upstream fix
* fix: lora kernel patch path for multimodal models
* fix: removed mllama from quarto
* run test for came optim on 2.6.0+
* fix fsdp2 patch and remove deprecated patch
* make sure to set sequence_parallel_degree for grpo
* Add SP test for GRPO
* add sp to grpo config for trainer
* use reward_funcs as kwarg to grpo trainer
* fix the comprehension for reward funcs
* reward funcs already passed in as args
* init sp_group right before training
* fix check for adding models to SP context
* make sure to pass args to super
* upgrade deepspeed
* use updated trl and add reasoning flags for vllm
* patch the worker
---------
Co-authored-by: NanoCode012 <nano@axolotl.ai>
* don't set peft_config on grpo to prevent double peft wrap
* remove overrides needed to support bug
* fix grpo tests
* require more CPU for multigpu to help with torch compile for vllm
* update doc and skip brittle grpo test
* fix the path to run the multigpu tests
* increase timeout, use LOC instead of NVL
* typo
* use hf cache from s3 backed cloudfront
* mark grpo as flaky test dues to vllm start
* update trl to 0.17.0
* grpo + vllm no longer supported with 2.5.1 due to vllm constraints
* disable VLLM_USE_V1 for ci
* imporve handle killing off of multiprocessing vllm service
* debug why this doesn't run in CI
* increase vllm wait time
* increase timeout to 5min
* upgrade to vllm 0.8.4
* dump out the vllm log for debugging
* use debug logging
* increase vllm start timeout
* use NVL instead
* disable torch compile cache
* revert some commented checks now that grpo tests are fixed
* increase vllm timeoout back to 5min
* batch api HF adapter for ring-flash-attn; cleanup and improvements
* update
* adding all batch ring-flash-attn methods via single adapter
* removing pad_to_sequence_len=False for now
* fix
* updating docs to include batch SP
* review comments
* fixes for batch API funcs, simplify
* fixes
* fix
* updates
* add batch_zigzag smoke test
* fixes for delinearization, and make qlora work with fsdp2
* Add back mistakenly removed lm_eval
* typo [skip ci]
* patch evals for torch.compile + fsdp2
* also check torch_compile w fsdp2
* lots of fixes for flex attn with llama4
* fix patch check and patch llama4 too
* attempt to make the patches stick
* use transformers 4.51.2
* update configs and README for llama4
* remove torch.compile for CI test
* cleanup any existing singletons
* set singleton cache to None instead of deleting
* use importlib reload with monkeypatch
* don't worry about transformers version, mark inputs with grads, fix regex
* make sure embeds aren't on cpu
* logging and mem improvements
* vllm version and add to docker, make sure to save processor on conversion
* fix ambiguous tensor bool check
* fix vllm to not use v1, upgrade hf transformers
* fix tests
* make flex_attn_compile_kwargs configurable, since this depends on model params
---------
Co-authored-by: Wing Lian <wing@axolotl.ai>
Co-authored-by: Salman Mohammadi <salman.mohammadi@outlook.com>
* [ci] make e2e tests a bit faster by reducing test split size
* use 10% split of alpaca dataset to speed up dataset loading/tokenization
* reduce gas 4->2 for most e2e tests
* increase val set size for packing
* llama4 support
* add xet support [skip ci]
* be flexible on transformers version and skip test on version
* don't use deepspeed for the fix_untrained_tokens test
* reordering to trigger torch 2.6.0 tests first
* slightly smaller train set
* use 4.51.0 for now
* remove stray print, add llama4 chat template to schema, bump peft to 0.15.1
* patches to make llama4 performant
* add preliminary fp8 support
* fsdp2 support
* use accelerate release 1.6.0
* allow 8bit optims with fsdp2
* liger + torch compile fix
* add fsdp2 e2e tests
* use transformers commit with fsdp2 support
* skip zero3 tests for this PR for now
* fix fsdp2 config for ci
* make sure both flex and flash attn work with fsdp2, skip fix untrained tokens
* okay, actually use fdsp2...
* more fixes to flex for fsdp2
* make sure to patch all the loaded models
* additional validation for fsdp2, bump dep versions
* make gemma3 work with packing
* multi-gpu e2e for ci
* update gemma3 model namespace to use mirror
* add gradient checkpointing to multigpu e2e ci
* update gemma3 examples for use_reentrant and fix ddp find unused params
* fix tests for gemma3
* fix import for test utils
* set correct train loss for gemma3 e2e
* add grpo scale_rewards config for trl#3135
* options to connect to vllm server directly w grpo trl#3094
* temperature support trl#3029
* sampling/generation kwargs for grpo trl#2989
* make vllm_enable_prefix_caching a config param trl#2900
* grpo multi-step optimizeations trl#2899
* remove overrides for grpo trainer
* bump trl to 0.16.0
* add cli to start vllm-serve via trl
* call the python module directly
* update to use vllm with 2.6.0 too now and call trl vllm serve from module
* vllm 0.8.1
* use python3
* use sys.executable
* remove context and wait for start
* fixes to make it actually work
* fixes so the grpo tests pass with new vllm paradigm
* explicit host/port and check in start vllm
* make sure that vllm doesn't hang by setting quiet so outouts go to dev null
* also bump bnb to latest release
* add option for wait from cli and nccl debugging for ci
* grpo + vllm test on separate devices for now
* make sure grpo + vllm tests runs single worker since pynccl comms would conflict
* fix cli
* remove wait and add caching for argilla dataset
* refactoring configs
* chore: lint
* add vllm config
* fixup vllm grpo args
* fix one more incorrect schema/config path
* fix another vlllm reference and increase timeout
* make the tests run a bit faster
* change mbsz back so it is correct for grpo
* another change mbsz back so it is correct for grpo
* fixing cli args
* nits
* adding docs
* docs
* include tensor parallel size for vllm in pydantic schema
* moving start_vllm, more docs
* limit output len for grpo vllm
* vllm enable_prefix_caching isn't a bool cli arg
* fix env ordering in tests and also use pid check when looking for vllm
---------
Co-authored-by: Salman Mohammadi <salman.mohammadi@outlook.com>
