Build the Qwen3 attention mask in the query dtype - #34
Merged
Conversation
Since #33 the Qwen3 encoder's additive mask goes through the fused SDPA kernel, whose CUDA memory-efficient backend requires the bias dtype to match the query; the mask was always float32 against bf16 q/k/v, so every FLUX.2 prompt encode on GPU failed with 'invalid dtype for bias'. Build the mask in the embedding dtype instead: -10000 fill (the dit_ltx23 convention; -3.4e38 overflows bfloat16 to -Inf) and the $neg()$add(1) chain, since the Ops dispatch materializes R scalars as Float tensors and promotes the result.
TroyHernandez
added a commit
that referenced
this pull request
Jul 20, 2026
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Every FLUX.2 prompt encode on CUDA failed with "invalid dtype for bias - should match query's dtype" since #33 routed the Qwen3 encoder's additive mask through the fused SDPA kernel, whose memory-efficient backend requires the bias dtype to match the query. The mask was always float32 against bf16 q/k/v.
The mask is now built in the embedding dtype. Two adjacent traps handled along the way:
1 - tensorOps dispatch materializes the scalar as a dimensioned Float tensor and promotes bf16 back to f32, so the pad term uses the$neg()$add(1)$mul()chain instead (upstream fix proposed in Preserve tensor dtype when the left operand is an R scalar mlverse/torch#1471).Verified with a kernel-level repro and a small-config encoder forward: CUDA bf16 and CPU f32 both pass with finite outputs. The LTX-2.3 transformer already casts its masks to the hidden-state dtype, so the stv path was unaffected.