Conversation
…ift causal-training gate - Metal Flash Attention calls now go through PyTorch's SDPA dispatcher (which current UMFA builds register on the MPS key) instead of calling extension autograd entrypoints directly, so fp32/int8/int4 profiles all use UMFA's registered mask-aware path. - Startup parity checks now cover fp16/bf16 forward and autograd plus causal forward against a CPU reference, and get_dispatch_stats() is required unconditionally since dispatcher routing depends on it. - Eligibility no longer rejects supported MPS masks, fp16/bf16 inputs, small verified shapes, or causal calls that require gradients — UMFA's causal backward passes exact dQ/dK/dV parity on both transposed masking codegen paths, so causal training dispatches through UMFA. - Remaining fallbacks: dropout and enable_gqa.
Rewrites the stale eligibility description (previously 'unmasked, at least four heads, sequence length 64+') across OPTIONS and the experimental UMFA guide, in all six languages: - Dispatcher-based routing with in-stream encoding: any head count (single-head works), any sequence length, bool/additive masks up to 4D, causal calls with and without gradients (causal backward passes exact gradient parity). Fallbacks are dropout and enable_gqa. - fp16/bf16 run native low-precision kernels with no fp32 promotion; mixed_precision bf16 no longer needs fp32 overrides, and the enable example drops the old fp32 pinning. - Mask support documented for the regular dispatcher (not just quantized), including the all-true fast path and in-stream expansion. - Dispatch counters: fp32_instream for unquantized runs, rope_instream for the fused entry point, quantized_autograd for the aliases. - New fused RoPE + SDPA section: rope_scaled_dot_product_attention covering the shared interleaved-pair convention of FLUX.1/FLUX.2/ Krea2/Z-Image, pair-duplicated [S,D]/[1,S,D]/[B,S,D] tables, strided views, fused autograd with inverse rotation on dQ/dK, and the measured 4.4 ms/layer gain at the Z-Image DiT shape.
metal-flash-attention: update api for new constants and masked attention API
…n for Flux1/Flux2, Z-Image and Krea
…tions metal-flash-attention: causal training ungated; fused RoPE + Attention for Flux1/Flux2, Z-Image and Krea
…ints Checkpoints tab "Run Inference" button to run inference pipeline from
wan 2.2: add AnimeGen T2V model flavour
…actor deepfloyd: refactor pipeline handling (#2604)
(#2812) fix vllm version for newer torch dependency
(#2607) Wan MoE validation improvements
…ference (#2605) SDXL refiner two-stage pipeline inference
refactor model-specific registry fields into the models subdir instead
(#2606) pixart two-stage pipeline inference support (v0.7 MoE variant)
docs: update SDXL quickstart validation notes
docs: update Wan quickstart validation notes
docs: update DeepFloyd validation notes
docs: update PixArt quickstart validation notes
(#2603) LTX-2.3 (and compatible) two-stage validation via upscaler 2nd stage
(#2819) fix nvidia dependency versions for cuda13
…-lora (#2820) validation adapter assignment into other (frozen/untrained) inference stages
remove vLLM from base dependencies; it is captioning only
Optionally use Qwen VL driven REPA instead of DINO
…onfig (#1814) validation image inputs without full dataset configuration
…ine correct class
bugfix: unload model button should remain visible after env switch or reload
update model metadata cache
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.
No description provided.