-
Notifications
You must be signed in to change notification settings - Fork 131
All issues
Issue creation is restricted in this repository
- #4938 · chuanqi129 opened
on Aug 13, 2026 8 - #5490 · torchxpubot opened
on Sep 22, 2026 38
Issues
is:issue state:open
is:issue state:open
Search results
[xpu-alignment] [26-09-25] XPU layer_norm returns non-zero on a constant input: vectorized WelfordCombine averages partial means instead of using Welford's increment
ai_generatedContent generated by an AI assistantContent generated by an AI assistantStatus: Open.#5535 In intel/torch-xpu-ops;- Status: Open.#5529 In intel/torch-xpu-ops;
- Status: Open.#5528 In intel/torch-xpu-ops;
- Status: Open.#5527 In intel/torch-xpu-ops;
- Status: Open.#5526 In intel/torch-xpu-ops;
[XPU] bf16
F.linearis not run-to-run reproducible by default for large-K / small-M shapes (CUDA is)Status: Open.#5524 In intel/torch-xpu-ops;[xpu-alignment] [26-09-24] XPU linspace/logspace compute the half and bfloat16 step in the output dtype, so results diverge from CPU
ai_generatedContent generated by an AI assistantContent generated by an AI assistantStatus: Open.#5523 In intel/torch-xpu-ops;FlopCounterMode reports 0 FLOPs for SDPA on XPU (
_scaled_dot_product_fused_attention_overrideablehas no flop formula)test: utUnit-test failure surfaceUnit-test failure surfaceStatus: Open.- Status: Open.#5515 In intel/torch-xpu-ops;
[OneDNN3.13][BMG] timm_models amp_fp16/training fbnetv3_b got fail_accuracy
AccuracyAccuracy-related failureAccuracy-related failuredependency component: oneDNNoneDNN primitive: matmul/conv/linear/quantizedoneDNN primitive: matmul/conv/linear/quantizedStatus: Open.- Status: Open.#5510 In intel/torch-xpu-ops;
- Status: Open.#5509 In intel/torch-xpu-ops;