Search results

[BMG][PyTorch] Detectron2 inductor got crash with Triton 3.8.0 on B60

tests: e2e

upstream: pytorch

Status: Open.

Bug

#7512 In intel/intel-xpu-backend-for-triton;

· mengfei25 opened on Jul 17, 2026

·  

[PT 2.14]

[BMG][PyTorch] Training performance regressions with Triton 3.8.0 vs 3.7.2 on B60 Linux

performance

tests: e2e

Status: Open.

Bug

#7511 In intel/intel-xpu-backend-for-triton;

· mengfei25 opened on Jul 17, 2026

·  

[PT 2.14]

[Bug] Tensor-descriptor uint8 load × bf16 scale → tl.dot produces incorrect results when BLOCK_SIZE_N>=64 and BLOCK_SIZE_M!=32

vllm

Status: Open.

Bug

#7510 In intel/intel-xpu-backend-for-triton;

· afierka-intel opened on Jul 17, 2026

[VLLM][FUSED_MOE] Add TritonGen operation for vector gather load for flexable return layout.

Status: Open.

#7509 In intel/intel-xpu-backend-for-triton;

· chengjunlu opened on Jul 17, 2026

·  

5. Ecosystem enabling and performance

[Windows] test_compile_only_* fail on Intel-only CI: missing ptxas-blackwell.exe / AMD toolchain

tests: ut

windows

Status: Open.

#7504 In intel/intel-xpu-backend-for-triton;

· kwasd opened on Jul 16, 2026

[Windows] test_chain_reduce[1-max-src_layout2-256-*] exceeds PTSS scratch space limit on A770/B580

hw: arc

hw: bmg

tests: ut

windows

Status: Open.

#7503 In intel/intel-xpu-backend-for-triton;

· kwasd opened on Jul 16, 2026

[BMG][Windows] timm_models_accuracy_inference_amp-bf16 pit_b_224 fail accuracy with triton 3.8.0

hw: bmg

tests: e2e

upstream: pytorch

windows

Status: Open.

Bug

#7496 In intel/intel-xpu-backend-for-triton;

· bjarzemb opened on Jul 16, 2026

[BMG][Windows] Performance regressions with triton 3.8.0

hw: bmg

tests: e2e

windows

Status: Open.

#7495 In intel/intel-xpu-backend-for-triton;

· bjarzemb opened on Jul 16, 2026

Regressions on "PyTorch Flex Attention, PVC (Max 1100), pinned"

tests: torchinductor

upstream: pytorch

Status: Open.

Bug

#7493 In intel/intel-xpu-backend-for-triton;

· anmyachev opened on Jul 16, 2026

Regressions on "PyTorch Flex Decoding, PVC (Max 1100), pinned"

tests: torchinductor

upstream: pytorch

Status: Open.

Bug

#7492 In intel/intel-xpu-backend-for-triton;

· anmyachev opened on Jul 16, 2026

[codegen] arith.truncf f32->f16 followed by arith.extf f16->f32 is folded to identity by SPIR-V/IGC, breaking numerical barrier

codegen: mlir

upstream: pytorch

Status: Open.

Bug

#7491 In intel/intel-xpu-backend-for-triton;

· ZhaoqiongZ opened on Jul 16, 2026

[A770][Windows] test_debug.py device_assert/sanitize_int_overflow subprocess checks fail: SIGABRT reported as Windows exit code 0xc0000409

hw: arc

tests: ut

windows

Status: Open.

#7481 In intel/intel-xpu-backend-for-triton;

· kwasd opened on Jul 15, 2026