Support tensor-valued _extra_state values
TransformerEngine uses the pytorch get/set_extra_state API to store FP8
layer config information as bytes Tensor in the _extra_state entry in
the state dict. With recent changes to from_pretrained, this
functionality has broken and loading a model that uses this API doesn't
appear to work. This PR fixes the save/load pretrained functions for
extra state entries that use a pytorch tensor, and adds a (currently
x-failing) test for a dictionary extra state.
Signed-off-by: Peter St. John <pstjohn@nvidia.com>
* Firstly: Better detection of when we're a custom class
* Trigger tests
* Let's break everything
* make fixup
* fix mistaken line doubling
* Let's try to get rid of it from config classes at least
* Let's try to get rid of it from config classes at least
* Fixup image processor
* no more circular import
* Let's go back to setting `_auto_class` again
* Let's go back to setting `_auto_class` again
* stash commit
* Revert the irrelevant changes until we figure out AutoConfig
* Change tests since we're breaking expectations
* make fixup
* do the same for all custom classes
* Cleanup for feature extractor tests
* Cleanup tokenization tests too
* typo
* Fix tokenizer tests
* make fixup
* fix image processor test
* make fixup
* Remove warning from register_for_auto_class
* Stop adding model info to auto map entirely
* Remove todo
* Remove the other todo
* Let's start slapping _auto_class on models why not
* Let's start slapping _auto_class on models why not
* Make sure the tests know what's up
* Make sure the tests know what's up
* Completely remove add_model_info_to_*
* Start adding _auto_class to models
* Start adding _auto_class to models
* Add a flaky decorator
* Add a flaky decorator and import
* stash commit
* More message cleanup
* make fixup
* fix indent
* Fix trust_remote_code prompts
* make fixup
* correct indentation
* Reincorporate changes into dynamic_module_utils
* Update call to trust_remote_code
* make fixup
* Fix video processors too
* Fix video processors too
* Remove is_flaky additions
* make fixup
* Get parallel loader working. Include tests.
* Update the tests for parallel loading
* Rename env variables.
* Add docs for parallel model weight loading.
* Touch up parallel model loading docs.
* Touch up parallel model loading docs again.
* Edit comment in test_modeling_utils_parallel_loading.py
* Make sure HF_PARALLEL_LOADING_WORKERS is spelled correctly in modeling_utils.py
* Correct times for parallelized loading, previous times were for a "hot" filesystem
* Update parallel model loading so the spawn method is encapsulated. DRY up the code by leveraging get_submodule.
* Update docs on model loading parallelism so that details on setting the multiprocessing start method are removed, now that the package handles this step internally.
* Fix style on model loading parallelism changes.
* Merge latest version of master's modeling_utils.
* Removed unused variable.
* Fix argument packing for the parallel loader.
* Fix state dict being undefined in the parallel model loader.
* Rename variables used in parallel model loading for clarity. Use get_module_from_name().
* Switch to the use of threads for parallel model loading.
* Update docs for parallel loading.
* Remove the use of json.loads when evaluating HF_ENABLE_PARALLEL_LOADING. Prefer simple casting.
* Move parallelized shard loading into its own function.
* Remove use of is_true(). Favor checking env var true values for HF_ENABLE_PARALLEL_LOADING.
* Update copyright to 2025 in readme for paralell model loading.
* Remove garbage collection line in load_shard_file, implicit garbage collection already occurs.
* Run formatter on modeling_utils.py
* Apply style fixes
* Delete tests/utils/test_modeling_utils_parallel_loading.py
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co>
* stash for now
* initial commit
* small updated
* up
* up
* works!
* nits and fixes
* don't loop too much
* finish working example
* update
* fix the small freeblocks issue
* feat: stream inputs to continuous batch
* fix: update attn from `eager` to `sdpa`
* refactor: fmt
* refactor: cleanup unnecessary code
* feat: add `update` fn to `PagedAttentionCache`
* feat: broken optimal block size computation
* fix: debugging invalid cache logic
* fix: attention mask
* refactor: use custom prompts for example
* feat: add streaming output
* fix: prefill split
refactor: add doc strings and unsound/redundant logic
fix: compute optimal blocks logic
* fix: send decoded tokens when `prefilling_split` -> `decoding`
* refactor: move logic to appropriate parent class
* fix: remove truncation as we split prefilling anyways
refactor: early return when we have enough selected requests
* feat: add paged attention forward
* push Ggraoh>
* add paged sdpa
* update
* btter mps defaults
* feat: add progress bar for `generate_batch`
* feat: add opentelemetry metrics (ttft + batch fill %age)
* feat: add tracing
* Add cuda graphs (#38059)
* draft cudagraphs addition
* nits
* styling
* update
* fix
* kinda draft of what it should look like
* fixes
* lol
* not sure why inf everywhere
* can generate but output is shit
* some fixes
* we should have a single device synch
* broken outputs but it does run
* refactor
* updates
* updates with some fixes
* fix mask causality
* another commit that casts after
* add error
* simplify example
* update
* updates
* revert llama changes
* fix merge conflicts
* fix: tracing and metrics
* my updates
* update script default values
* fix block allocation issue
* fix prefill split attnetion mask
* no bugs
* add paged eager
* fix
* update
* style
* feat: add pytorch traces
* fix
* fix
* refactor: remove pytorch profiler data
* style
* nits
* cleanup
* draft test file
* fix
* fix
* fix paged and graphs
* small renamings
* cleanups and push
* refactor: move tracing and metrics logic to utils
* refactor: trace more blocks of code
* nits
* nits
* update
* to profile or not to profile
* refactor: create new output object
* causal by default
* cleanup but generations are still off for IDK what reason
* simplifications but not running still
* this does work.
* small quality of life updates
* nits
* updaet
* fix the scheduler
* fix warning
* ol
* fully fixed
* nits
* different generation parameters
* nice
* just style
* feat: add cache memory usage
* feat: add kv cache free memory
* feat: add active/waiting count & req latency
* do the sampling
* fix: synchronize CUDA only if available and improve error handling in ContinuousBatchingManager
* fix on mps
* feat: add dashboard & histogram buckets
* perf: improve waiting reqs data structures
* attempt to compile, but we should only do it on mps AFAIK
* feat: decouple scheduling logic
* just a draft
* c;eanup and fixup
* optional
* style
* update
* update
* remove the draft documentation
* fix import as well
* update
* fix the test
* style doomed
---------
Co-authored-by: Luc Georges <luc.sydney.georges@gmail.com>
* starting attn refactor for encoder decoder models via bart (eager + sdpa)
* flash attention works, remove unnecessary code
* flex attention support for bart!, gotta check if the renaming is not too aggressive
* some comments
* skip flex grad test for standalone as done with the other test
* revert flex attn rename (for now), sdpa simplify, and todos
* more todos
* refactor mask creation for reuse
* modular attempt at biogpt
* first batch of other models
* fix attn dropout
* fix autoformer copies
* hubert
* another batch of models
* copies/style + last round of bart models --> whisper next?
* remove unnecessary _reshape function and remove copy to whisper
* add skip for decoder-only models out of enc-dec (same as in bart)
* bring back licences
* remove comment, added to pr read instead
* mostly docs
* disable sew flex attn as it's unclear attn mask for now
* oops
* test fixes for enc-dec
* torch fx fixes + try at flex attn
* skip on mbart
* some more fixes
* musicgen skip / delete old attn class logic + sdpa compose compile skip
* disable flex attn for musicgen, not worth the effort
* more fixes and style
* flex attention test for dropout and encoder decoder that dont have main input names
* informer fixes
* the weirdest thing I've encountered yet...
* style
* remove empty tensor attempt, found core root in previous commits
* disable time series due to tests being very text centric on inputs
* add speech to text to be ignoring the other attns, also due to tests
* update docs
* remaining issues resolved ?
* update docs for current state --> nllb moe and pegasus x sdpa is questionable :D
* some models have not set the is_causal flag...
* change dtype in softmax tol old behaviour + some modular fixes
* I hate it but it is what it is
* fixes from main for bart
* forgot this one
* some model fixes
* style
* current status
* marian works now
* fixing some copies
* some copy fixes + time series x informer
* last models possibly and fixes on style/copies
* some post merge fixes
* more fixes
* make attention interface callable and move warnings there
* style lol
* add comment to "unsupported"
* remove callable interface and change interface warnings + some copies
* fix
* ternary is ugly af, make it simpler
* how did that happen
* fix flex attn test
* failing the test
* no more fallback! fixing copies next
* style + attn fixed
* fixing copies and mask creation
* wrong copy
* fixup tests and disable flex attn for now
* fixup last tests?
* accept custom device_mesh
* fix device_map
* assert that num_heads % tp_size == 0
* todo.
* ReplicateParallel
* handle tied weights
* handle dtensor in save_pretrained with safe_serialization
* tp test works
* doesnt work
* fix shard_and_distribute_module's rank should be local_rank
* tp=4 is correct
* dp+tp is broken
* todo allreduce with dtensors on another dim is annoying
* workaround to sync dp grads when using dtensors
* loading a checkpoint works
* wandb and compare losses with different tp/dp
* cleaning
* cleaning
* .
* .
* logs
* CP2 DP2 no mask works after commenting attn_mask and is_causal from scaled_dot_product_attention
* DP=2 TP=2 now works even with tied embeddings
* model.parameters() and model.module.parameters() are empty..
* reformat sanity_check_tensor_sync
* set atol=1e-4 for CP to pass
* try populate _parameters from named_modules
* refactors
TP2 DP2 works
CP2 DP2 works
* is_causal=True and pack sequences, no attn mask, and preshuffle dataset
* fix packing
* CP=4 doesn't work
* fix labels and position_ids for CP
* DP CP works with transformers 🥳🥳🥳
* refactor
* add example cp
* fixup
* revert sdpa changes
* example cleared
* add CP, DP to the mesh init
* nit
* clean
* use `ALL_PARALLEL_STYLES`
* style
* FSDP works
* log on 1 rank
* .
* fix?
* FSDP1 also has .parameters() bug
* reported gradnorm when using FSDP1 is wrong, but loss is correct so it's okay
* .
* style and fixup
* move stuff around
* fix tests
* style
* let's make it a check
* add missing licences
* warning should be an info
* tp plan should not be NONE
* test all
* god damn it
* test all
---------
Co-authored-by: nouamanetazi <nouamane98@gmail.com>
* accept custom device_mesh
* fix device_map
* assert that num_heads % tp_size == 0
* todo.
* ReplicateParallel
* handle tied weights
* handle dtensor in save_pretrained with safe_serialization
* tp test works
* doesnt work
* fix shard_and_distribute_module's rank should be local_rank
* tp=4 is correct
* dp+tp is broken
* todo allreduce with dtensors on another dim is annoying
* workaround to sync dp grads when using dtensors
* loading a checkpoint works
* wandb and compare losses with different tp/dp
* cleaning
* cleaning
* .
* .
* logs
* CP2 DP2 no mask works after commenting attn_mask and is_causal from scaled_dot_product_attention
* DP=2 TP=2 now works even with tied embeddings
* model.parameters() and model.module.parameters() are empty..
* reformat sanity_check_tensor_sync
* set atol=1e-4 for CP to pass
* try populate _parameters from named_modules
* refactors
TP2 DP2 works
CP2 DP2 works
* is_causal=True and pack sequences, no attn mask, and preshuffle dataset
* fix packing
* CP=4 doesn't work
* fix labels and position_ids for CP
* DP CP works with transformers 🥳🥳🥳
* refactor
* add example cp
* fixup
* revert sdpa changes
* example cleared
* add CP, DP to the mesh init
* nit
* clean
* use `ALL_PARALLEL_STYLES`
* style
* FSDP works
* log on 1 rank
* .
* fix?
* FSDP1 also has .parameters() bug
* reported gradnorm when using FSDP1 is wrong, but loss is correct so it's okay
* .
* style and fixup
* move stuff around
* fix tests
* style
* let's make it a check
* warning should be an info
---------
Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>
* mvp
* remove trust_remote_code
* generate_from_hub
* handle requirements; docs
* english
* doc PR suggestions
* Apply suggestions from code review
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* changed remote code path to generate/generate.py
* model repo has custom generate -> override base generate
* check for proper inheritance
* some doc updates (missing: tag-related docs)
* update docs to model repo
* nit
* nit
* nits
* Update src/transformers/dynamic_module_utils.py
* Apply suggestions from code review
* Update docs/source/en/generation_strategies.md
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* trust remote code is required
* use new import utils for requirements version parsing
* use org examples
* add tests
* Apply suggestions from code review
Co-authored-by: Manuel de Prada Corral <6536835+manueldeprada@users.noreply.github.com>
* ascii file structure; tag instructions on readme.md
---------
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Manuel de Prada Corral <6536835+manueldeprada@users.noreply.github.com>
* update models
* why rename
* return attn weights when sdpa
* fixes
* fix attn implementation composite
* fix moshi
* add message
* add typings
* use explicitly all flags for each attn type
* fix some tests
* import what is needed
* kosmos on main has ew attention already, yay
* new models in main, run fixup
* won't fix kosmos yet
* fix-copies
* clean up after rebasing
* fix tests
* style
* dont cast attns to fp32
* did we update ruff? oke, let's just do what it asks
* fix pixtral after rebase
* i guessreverted all CdGen classes
* style
* llava onevision
* fix copies
* fix some tests
* some more tests
* dump
* skip these
* nevermind, i am dumb
* revert fix not needed
* fixup
* fixup
* another fixup
* more fixup to make ci finally happy
* fixup after rebasing
* fix qwen tests
* add internVL + typos here and there
* image token index -> id
* style
* fix init weights
* revert blip-2 not supported
* address comments
* fix copies
* revert blip2 test file as well
* as discussed internally, revert back CdGen models
* fix some tests
* fix more tests for compile
* CI red
* fix copies
* enumerate explicitly allowed models
* address comments
* fix tests
* fixup
* style again
* add tests for new model class
* another fixup ( x _ x )
* [fixup] unused attributes can be removed post-deprecation
* skip compilation on cpu offload
* add test
* better logic
* docstring
* boolean logic
* add disk offload check
* warn users if compilation options are set but compilation doesn happen
* fix test
---------
Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
