[qwen3.5] fix DTensor TP vision position caches - #3899
Merged
Conversation
[ghstack-poisoned]
meta-cla
Bot
added
the
CLA Signed
pianpwk
marked this pull request as ready for review
July 10, 2026 18:09
pianpwk
requested review from
fegin,
tianyu-l,
wconstab and
wwwjn
as code owners
July 10, 2026 18:09
tianyu-l
approved these changes
Jul 10, 2026
tianyu-l
left a comment
tianyu-l
left a comment
Contributor
There was a problem hiding this comment.
pls make sure numerics doesn't change
jemitche1
pushed a commit
to jemitche1/torchtitan
that referenced
this pull request
Jul 15, 2026
Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.12.0) (oldest at bottom): * pytorch#3895 * __->__ pytorch#3899 Main TP was failing with `FLA_TILELANG=0 MODULE=qwen3_5 CONFIG=qwen35_debugmodel NGPU=2 ./run_train.sh --parallelism.tensor_parallel_degree 2 --parallelism.data_parallel_shard_degree 1 --debug.seed 42 --debug.deterministic --training.steps 1` ``` [rank0]:[rank0]: File "/data/users/pianpwk/full_spmd_types_torchtitan/tt4_spmd_types/torchtitan/torchtitan/models/qwen3_5/vision_encoder.py", line 91, in _compute_learned_pos_embeds [rank0]:[rank0]: pos_hw = F.interpolate( [rank0]:[rank0]: ^^^^^^^^^^^^^^ [rank0]:[rank0]: File "/data/users/pianpwk/full_spmd_types_torchtitan/pytorch/torch/nn/functional.py", line 5273, in interpolate [rank0]:[rank0]: )._upsample_linear_vec(input, output_size, align_corners, scale_factors) [rank0]:[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank0]:[rank0]: File "/data/users/pianpwk/full_spmd_types_torchtitan/pytorch/torch/_decomp/decompositions.py", line 4507, in _upsample_linear_vec [rank0]:[rank0]: return _upsample_linear(input, osize, align_corners, scales) [rank0]:[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank0]:[rank0]: File "/data/users/pianpwk/full_spmd_types_torchtitan/pytorch/torch/_decomp/decompositions.py", line 91, in inner [rank0]:[rank0]: r = f(*tree_map(increase_prec, args), **tree_map(increase_prec, kwargs)) [rank0]:[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank0]:[rank0]: File "/data/users/pianpwk/full_spmd_types_torchtitan/pytorch/torch/_decomp/decompositions.py", line 4627, in _upsample_linear [rank0]:[rank0]: v = aten._unsafe_index(input, idx) [rank0]:[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank0]:[rank0]: File "/data/users/pianpwk/full_spmd_types_torchtitan/pytorch/torch/_ops.py", line 1350, in __call__ [rank0]:[rank0]: return self._op(*args, **kwargs) [rank0]:[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^ [rank0]:[rank0]: RuntimeError: aten._unsafe_index.Tensor got mixed torch.Tensor and DTensor, need to convert all torch.Tensor to DTensor before calling distributed operators! ``` - DTensor decomposition goes to unsafe_index which creates local plain tensors - fixes by local-mapping interpolate - upcoming spmd_types PR would parallelize inv_freq, get ahead of this by also wrapping positions as DTensors - previously that portion was done in plain tensor
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
2 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.
Stack from ghstack (oldest at bottom):
Main TP was failing with
FLA_TILELANG=0 MODULE=qwen3_5 CONFIG=qwen35_debugmodel NGPU=2 ./run_train.sh --parallelism.tensor_parallel_degree 2 --parallelism.data_parallel_shard_degree 1 --debug.seed 42 --debug.deterministic --training.steps 1