# triton-lang/triton v3.9.0: Triton 3.9.0 Release Notes — retained record

> Retained github record from github:triton-lang/triton.

Canonical: https://highsignal.app/data/records/88407ce91ee2a3e8

[← sources](https://highsignal.app/data)


source: [github](https://highsignal.app/data/github) (github:triton-lang/triton)  published 2026-10-27 18:56 ingested 2026-10-28 02:33

[original source ↗](https://github.com/triton-lang/triton/releases/tag/v3.9.0)

## extract

# Triton 3.9.0 Release Notes

## Table of Contents

- [Dialect & Frontend](https://highsignal.app/data/records/88407ce91ee2a3e8#dialect--frontend-39)
- [Backend & Compiler](https://highsignal.app/data/records/88407ce91ee2a3e8#backend--compiler-39)
- [AMD/HIP Backend](https://highsignal.app/data/records/88407ce91ee2a3e8#amdhip-backend-39)
- [NVIDIA Backend](https://highsignal.app/data/records/88407ce91ee2a3e8#nvidia-backend-39)
- [Gluon & Layout Improvements](https://highsignal.app/data/records/88407ce91ee2a3e8#gluon--layout-improvements-39)
- [Kernels & Benchmarks](https://highsignal.app/data/records/88407ce91ee2a3e8#kernels--benchmarks-39)
- [Proton Profiling](https://highsignal.app/data/records/88407ce91ee2a3e8#proton-profiling-39)
- [Testing & CI](https://highsignal.app/data/records/88407ce91ee2a3e8#testing--ci-39)
- [Build & Infrastructure](https://highsignal.app/data/records/88407ce91ee2a3e8#build--infrastructure-39)
- [Documentation](https://highsignal.app/data/records/88407ce91ee2a3e8#documentation-39)
- [Deprecations](https://highsignal.app/data/records/88407ce91ee2a3e8#deprecations-39)
- [Breaking Changes](https://highsignal.app/data/records/88407ce91ee2a3e8#breaking-changes-39)
- [Contributors](https://highsignal.app/data/records/88407ce91ee2a3e8#contributors-39)

---

## Dialect & Frontend<a id="dialect--frontend-39"></a>

### New Features

- **Atomic loads and stores:** Added `tl.atomic_load` and `tl.atomic_store`, also available in Gluon and the interpreter. Loads default to `acquire` and stores default to `release` ([#11591](https://github.com/triton-lang/triton/pull/11591), [#11650](https://github.com/triton-lang/triton/pull/11650), [#11676](https://github.com/triton-lang/triton/pull/11676), [#11821](https://github.com/triton-lang/triton/pull/11821))
- **`tl.atomic_poll`:** Added `tl.atomic_poll`, which waits until the value at a pointer equals an expected value and returns `False` on timeout ([#10723](https://github.com/triton-lang/triton/pull/10723), [#10755](https://github.com/triton-lang/triton/pull/10755), [#11540](https://github.com/triton-lang/triton/pull/11540))
- **Approximate division:** Added `tl.fdiv(..., approx=True)` for fast FP32 division ([#11900](https://github.com/triton-lang/triton/pull/11900))
- **Language syntax:** Kernels support chained comparisons such as `0 <= x <= 8`, keyword-only parameters, tuple unpacking in list comprehensions, and operator methods on aggregate types ([#12002](https://github.com/triton-lang/triton/pull/12002), [#11418](https://github.com/triton-lang/triton/pull/11418), [#11076](https://github.com/triton-lang/triton/pull/11076), [#10990](https://github.com/triton-lang/triton/pull/10990))
- **Integer `tl.clamp`:** `tl.clamp` accepts integer tensors ([#12005](https://github.com/triton-lang/triton/pull/12005))

### Improvements

- **Earlier errors:** Mismatched tuple unpacking, invalid `tl.topk` `k` values, `tl.cat` shape mismatches, and non-32-bit `tl.histogram` inputs now fail at compile time ([#11053](https://github.com/triton-lang/triton/pull/11053), [#11074](https://github.com/triton-lang/triton/pull/11074), [#11353](https://github.com/triton-lang/triton/pull/11353), [#10891](https://github.com/triton-lang/triton/pull/10891))
- **Hints through helpers:** Alignment and contiguity hints on values returned from inlined helpers are now kept ([#11150](https://github.com/triton-lang/triton/pull/11150))
- **Compile time:** Deeply nested loops compile about 10x faster at nesting depth 8 ([#11369](https://github.com/triton-lang/triton/pull/11369))

### Bug Fixes

- **Wrong results:** Fixed `tl.softmax` with `dim` other than 0, integer stores to boolean pointers, list-comprehension `if` filters, negative-zero constants, and stale compilations when a captured constexpr shares a global's name ([#11409](https://github.com/triton-lang/triton/pull/11409), [#11853](https://github.com/triton-lang/triton/pull/11853), [#10888](https://github.com/triton-lang/triton/pull/10888), [#11426](https://github.com/triton-lang/triton/pull/11426), [#11224](https://github.com/triton-lang/triton/pull/11224))
- **E8M0 scale decoding:** E8M0 `0x00` now decodes to `2^-127` and `0xFF` to NaN (previously zero and infinity), which changes results for kernels that use those values ([#11624](https://github.com/triton-lang/triton/pull/11624))
- **Forked processes:** A forked child process no longer unloads GPU modules that belong to its parent ([#11128](https://github.com/triton-lang/triton/pull/11128))
- **Interpreter:** `TRITON_INTERPRET=1` now matches compiled kernels for integer overflow, `tl.fma`, `tl.umulhi`, and `tl.histogram` ([#10923](https://github.com/triton-lang/triton/pull/10923), [#11194](https://github.com/triton-lang/triton/pull/11194), [#11210](https://github.com/triton-lang/triton/pull/11210), [#11390](https://github.com/triton-lang/triton/pull/11390), [#11424](https://github.com/triton-lang/triton/pull/11424), [#11531](https://github.com/triton-lang/triton/pull/11531), [#11641](https://github.com/triton-lang/triton/pull/11641), [#11403](https://github.com/triton-lang/triton/pull/11403), [#11218](https://github.com/triton-lang/triton/pull/11218), [#11751](https://github.com/triton-lang/triton/pull/11751))

---

## Backend & Compiler<a id="backend--compiler-39"></a>

### LLVM Updates

- **LLVM pin:** Updated LLVM to llvm/llvm-project@b010a18d. AMD code generation now uses its own LLVM pin; see [AMD/HIP Backend](https://highsignal.app/data/records/88407ce91ee2a3e8#amdhip-backend-39) ([#10739](https://github.com/triton-lang/triton/pull/10739), [#11163](https://github.com/triton-lang/triton/pull/11163), [#11485](https://github.com/triton-lang/triton/pull/11485))

### Multi-CTA / Multicast / TMA

- **Multi-CTA kernels:** Fixed missing cluster synchronization and proxy fences. Multi-CTA kernels now support local gather from subslices, `tl.histogram`, and ordered atomics in warp-specialized kernels ([#10656](https://github.com/triton-lang/triton/pull/10656), [#10735](https://github.com/triton-lang/triton/pull/10735), [#10843](https://github.com/triton-lang/triton/pull/10843), [#10975](https://github.com/triton-lang/triton/pull/10975), [#10559](https://github.com/triton-lang/triton/pull/10559), [#11775](https://github.com/triton-lang/triton/pull/11775), [#10914](https://github.com/triton-lang/triton/pull/10914))
- **Unaligned TMA:** Fixed wrong results when per-stage TMA buffers are not 128-byte aligned ([#10898](https://github.com/triton-lang/triton/pull/10898))

### Code Generation & Analysis

- **Atomic ordering:** Atomics with `acquire`, `release`, or `acq_rel` semantics now insert CTA-wide barriers so the ordering covers the whole block. Read-modify-write and CAS atomics default to `acq_rel`; pass `sem="relaxed"` to avoid the barriers ([#10816](https://github.com/triton-lang/triton/pull/10816), [#11022](https://github.com/triton-lang/triton/pull/11022))
- **Barrier insertion:** Barrier insertion was rewritten as a standalone pass that adds missing barriers and fences and removes redundant ones ([#11389](https://github.com/triton-lang/triton/pull/11389), [#11391](https://github.com/triton-lang/triton/pull/11391), [#11396](https://github.com/triton-lang/triton/pull/11396), [#11401](https://github.com/triton-lang/triton/pull/11401), [#11411](https://github.com/triton-lang/triton/pull/11411), [#11414](https://github.com/triton-lang/triton/pull/11414), [#11431](https://github.com/triton-lang/triton/pull/11431), [#11462](https://github.com/triton-lang/triton/pull/11462), [#11468](https://github.com/triton-lang/triton/pull/11468), [#11479](https://github.com/triton-lang/triton/pull/11479), [#10087](https://github.com/triton-lang/triton/pull/10087), [#11324](https://github.com/triton-lang/triton/pull/11324), [#11464](https://github.com/triton-lang/triton/pull/11464), [#11499](https://github.com/triton-lang/triton/pull/11499), [#11597](https://github.com/triton-lang/triton/pull/11597))
- **Batched matmul:** On NVIDIA, batched MMAv2 dots are up to 40x faster ([#11822](https://github.com/triton-lang/triton/pull/11822))
- **Miscompile fixes:** Fixed wrong results in nested-loop flattening, reductions, scans, `tl.histogram`, gathers, axis analysis, and optimizations that moved or duplicated loads ([#11521](https://github.com/triton-lang/triton/pull/11521), [#11688](https://github.com/triton-lang/triton/pull/11688), [#11617](https://github.com/triton-lang/triton/pull/11617), [#11621](https://github.com/triton-lang/triton/pull/11621), [#10744](https://github.com/triton-lang/triton/pull/10744), [#11186](https://github.com/triton-lang/triton/pull/11186), [#11269](https://github.com/triton-lang/triton/pull/11269), [#11682](https://github.com/triton-lang/triton/pull/11682), [#11842](https://github.com/triton-lang/triton/pull/11842), [#11637](https://github.com/triton-lang/triton/pull/11637), [#11981](https://github.com/triton-lang/triton/pull/11981), [#11990](https://github.com/triton-lang/triton/pull/11990), [#11943](https://github.com/triton-lang/triton/pull/11943), [#12011](https://github.com/triton-lang/triton/pull/12011), [#11767](https://github.com/triton-lang/triton/pull/11767), [#10838](https://github.com/triton-lang/triton/pull/10838), [#11546](https://github.com/triton-lang/triton/pull/11546), [#11548](https://github.com/triton-lang/triton/pull/11548), [#11337](https://github.com/triton-lang/triton/pull/11337), [#11314](https://github.com/triton-lang/triton/pull/11314), [#11906](https://github.com/triton-lang/triton/pull/11906), [#11975](https://github.com/triton-lang/triton/pull/11975), [#11240](https://github.com/triton-lang/triton/pull/11240))

### Sanitizers & Debugging

- **ConSan:** ConSan catches more races and reports fewer false positives. It compiles and runs about 2.5x faster on one reported workload ([#10668](https://github.com/triton-lang/triton/pull/10668), [#10736](https://github.com/triton-lang/triton/pull/10736), [#10784](https://github.com/triton-lang/triton/pull/10784), [#10864](https://github.com/triton-lang/triton/pull/10864), [#10948](https://github.com/triton-lang/triton/pull/10948), [#11045](https://github.com/triton-lang/triton/pull/11045), [#11010](https://github.com/triton-lang/triton/pull/11010), [#11162](https://github.com/triton-lang/triton/pull/11162), [#11216](https://github.com/triton-lang/triton/pull/11216), [#11244](https://github.com/triton-lang/triton/pull/11244), [#11310](https://github.com/triton-lang/triton/pull/11310), [#11200](https://github.com/triton-lang/triton/pull/11200), [#11202](https://github.com/triton-lang/triton/pull/11202), [#11175](https://github.com/triton-lang/triton/pull/11175), [#11268](https://github.com/triton-lang/triton/pull/11268), [#11436](https://github.com/triton-lang/triton/pull/11436), [#11459](https://github.com/triton-lang/triton/pull/11459), [#11626](https://github.com/triton-lang/triton/pull/11626), [#11852](https://github.com/triton-lang/triton/pull/11852), [#11857](https://github.com/triton-lang/triton/pull/11857), [#11138](https://github.com/triton-lang/triton/pull/11138), [#11446](https://github.com/triton-lang/triton/pull/11446), [#11447](https://github.com/triton-lang/triton/pull/11447), [#11448](https://github.com/triton-lang/triton/pull/11448), [#11449](https://github.com/triton-lang/triton/pull/11449), [#11450](https://github.com/triton-lang/triton/pull/11450), [#11451](https://github.com/triton-lang/triton/pull/11451), [#11452](https://github.com/triton-lang/triton/pull/11452), [#11642](https://github.com/triton-lang/triton/pull/11642))
- **GSan:** GSan supports programmatic dependent launch, CUDA graphs built with explicit dependencies, and multi-CTA kernels, and runs up to 3.8x faster on MoE ([#10677](https://github.com/triton-lang/triton/pull/10677), [#10944](https://github.com/triton-lang/triton/pull/10944), [#11020](https://github.com/triton-lang/triton/pull/11020), [#11135](https://github.com/triton-lang/triton/pull/11135), [#11199](https://github.com/triton-lang/triton/pull/11199), [#11303](https://github.com/triton-lang/triton/pull/11303), [#11172](https://github.com/triton-lang/triton/pull/11172), [#11187](https://github.com/triton-lang/triton/pull/11187), [#11297](https://github.com/triton-lang/triton/pull/11297), [#11351](https://github.com/triton-lang/triton/pull/11351), [#12023](https://github.com/triton-lang/triton/pull/12023))
- **FpSan:** FpSan supports `tl.clamp`, pure inline assembly, and Gluon embedding ops. Narrowing casts now fold discarded bits by default; set `TRITON_FPSAN_HOMOMORPHIC_CASTS=1` to keep truncation ([#10707](https://github.com/triton-lang/triton/pull/10707), [#10627](https://github.com/triton-lang/triton/pull/10627), [#10908](https://github.com/triton-lang/triton/pull/10908), [#10834](https://github.com/triton-lang/triton/pull/10834), [#11005](https://github.com/triton-lang/triton/pull/11005), [#11083](https://github.com/triton-lang/triton/pull/11083), [#11101](https://github.com/triton-lang/triton/pull/11101), [#11159](https://github.com/triton-lang/triton/pull/11159), [#11170](https://github.com/triton-lang/triton/pull/11170), [#11489](https://github.com/triton-lang/triton/pull/11489), [#11627](https://github.com/triton-lang/triton/pull/11627), [#11665](https://github.com/triton-lang/triton/pull/11665), [#11716](https://github.com/triton-lang/triton/pull/11716), [#11974](https://github.com/triton-lang/triton/pull/11974))

### Extensions & Tooling

- **Loading extensions:** Extensions load at runtime through `extend_with` and `extend_dialects_with`, so a package can register itself on import ([#10775](https://github.com/triton-lang/triton/pull/10775))
- **Out-of-tree layouts:** Out-of-tree MMA and shared encodings work through `MmaEncodingTrait` and `SharedEncodingTrait` ([#11025](https://github.com/triton-lang/triton/pull/11025), [#11764](https://github.com/triton-lang/triton/pull/11764))

---

## AMD/HIP Backend<a id="amdhip-backend-39"></a>

### gfx1250 / CDNA 5

- **GFX1250-strict:** Added initial support for the `gfx1250-strict` target ([#11945](https://github.com/triton-lang/triton/pull/11945), [#12134](https://github.com/triton-lang/triton/pull/12134))
- **Tensor Data Movement:** Added Gluon `tdm.async_load_fused` and pipelining of in-loop descriptor stores and scatters. Fixed rank-reducing loads, multi-CTA gather/scatter, and size-1 block dimensions. Non-zero padding (`padding_option="nan"`) now raises an error ([#10628](https://github.com/triton-lang/triton/pull/10628), [#10564](https://github.com/triton-lang/triton/pull/10564), [#11077](https://github.com/triton-lang/triton/pull/11077), [#11833](https://github.com/triton-lang/triton/pull/11833), [#11963](https://github.com/triton-lang/triton/pull/11963), [#12331](https://github.com/triton-lang/triton/pull/12331), [#12441](https://github.com/triton-lang/triton/pull/12441), [#10870](https://github.com/triton-lang/triton/pull/10870))
- **Scaled conversions:** Gluon adds `scaled_downcast` to MXFP4/MXFP8 on CDNA4 and CDNA5, and `scaled_upcast` accepts compact scales ([#11377](https://github.com/triton-lang/triton/pull/11377), [#11442](https://github.com/triton-lang/triton/pull/11442), [#10979](https://github.com/triton-lang/triton/pull/10979))

### Other Targets

- **RDNA:** Added RDNA 4m (`gfx117x`) targets and RDNA3 buffer atomics. bf16 multiplication now rounds to nearest even, and `dot_scaled` rounds its bf16 downcast toward zero ([#10878](https://github.com/triton-lang/triton/pull/10878), [#11474](https://github.com/triton-lang/triton/pull/11474), [#11535](https://github.com/triton-lang/triton/pull/11535), [#11227](https://github.com/triton-lang/triton/pull/11227), [#11629](https://github.com/triton-lang/triton/pull/11629))

### Code Generation and Runtime

- **Separately pinned LLVM:** AMD code generation uses its own LLVM pin (triton-lang/llvm-project@6bc4aaf6). For a custom build, set `TRITON_AMD_CODEGEN_PATH` to a prebuilt `libtriton_amd_codegen` library ([#11485](https://github.com/triton-lang/triton/pull/11485), [#11717](https://github.com/triton-lang/triton/pull/11717), [#11745](https://github.com/triton-lang/triton/pull/11745), [#11912](https://github.com/triton-lang/triton/pull/11912), [#11925](https://github.com/triton-lang/triton/pull/11925), [#11916](https://github.com/triton-lang/triton/pull/11916), [#12003](https://github.com/triton-lang/triton/pull/12003), [#12016](https://github.com/triton-lang/triton/pull/12016))
- **TheRock:** Triton finds and prefers ROCm from TheRock Python wheels (`rocm_sdk`) unless `TRITON_*` paths are set. It raises an error if a different HIP runtime is already loaded ([#11028](https://github.com/triton-lang/triton/pull/11028), [#11398](https://github.com/triton-lang/triton/pull/11398), [#11492](https://github.com/triton-lang/triton/pull/11492))

### Bug Fixes

- **Incorrect results:** Fixed NaN handling in FP8 FNUZ upcasts and CDNA3 TF32 `tl.dot`, CDNA barriers emitted on RDNA, buffer-atomic fence ordering, and miscompiles in range analysis and canonicalization ([#10747](https://github.com/triton-lang/triton/pull/10747), [#11783](https://github.com/triton-lang/triton/pull/11783), [#10906](https://github.com/triton-lang/triton/pull/10906), [#11081](https://github.com/triton-lang/triton/pull/11081), [#11313](https://github.com/triton-lang/triton/pull/11313), [#11246](https://github.com/triton-lang/triton/pull/11246), [#11239](https://github.com/triton-lang/triton/pull/11239))
- **Hangs and races:** `tl.load(..., volatile=True)` keeps its volatility, which fixes hanging polling loops. Also fixed LDS races, wrong shared-memory addresses on RDNA3/3.5, out-of-bounds XF32 MFMA access, and a deadlock when loading HIP modules in parallel ([#11223](https://github.com/triton-lang/triton/pull/11223), [#11169](https://github.com/triton-lang/triton/pull/11169), [#11765](https://github.com/triton-lang/triton/pull/11765), [#11719](https://github.com/triton-lang/triton/pull/11719), [#11651](https://github.com/triton-lang/triton/pull/11651), [#11566](https://github.com/triton-lang/triton/pull/11566), [#11727](https://github.com/triton-lang/triton/pull/11727), [#11265](https://github.com/triton-lang/triton/pull/11265))
- **Compile cache:** Changing an external library now invalidates the compile cache ([#11884](https://github.com/triton-lang/triton/pull/11884))

---

## NVIDIA Backend<a id="nvidia-backend-39"></a>

### Matrix Instructions

- **MMA performance:** TCGen5 MMA reuses A operands across N tiles, and FP4×FP4 kernels on SM103 are 6-15% faster ([#11345](https://github.com/triton-lang/triton/pull/11345), [#11438](https://github.com/triton-lang/triton/pull/11438))
- **SM120:** Added batched (rank-3) scaled-dot scale layouts ([#11262](https://github.com/triton-lang/triton/pull/11262))
- **MMA fixes:** Fixed wrong results for FP4 MMA with mixed packing, Blackwell MMAs that share an accumulator across pipeline stages, small-K MMAv2 dots with upcast operands, and Hopper `warp_group_dot_wait` with `num_warps > 4` ([#11151](https://github.com/triton-lang/triton/pull/11151), [#11245](https://github.com/triton-lang/triton/pull/11245), [#10988](https://github.com/triton-lang/triton/pull/10988), [#11691](https://github.com/triton-lang/triton/pull/11691), [#11056](https://github.com/triton-lang/triton/pull/11056))

### TMA / Cluster Launch Control

- **CLC for Triton kernels:** On SM100 and newer, the `clc=True` launch option turns a one-tile-per-program kernel into a persistent Cluster Launch Control loop, including with `warp_specialize=True` ([#11124](https://github.com/triton-lang/triton/pull/11124), [#11116](https://github.com/triton-lang/triton/pull/11116))
- **TMA fixes:** Fixed TMA on shared-memory subviews, multicast im2col, device-side tensor-map updates, and loops that create tensor descriptors ([#11064](https://github.com/triton-lang/triton/pull/11064), [#11136](https://github.com/triton-lang/triton/pull/11136), [#11480](https://github.com/triton-lang/triton/pull/11480), [#10281](https://github.com/triton-lang/triton/pull/10281))

### Code Generation

- **ptxas:** Hopper (sm_90) and newer use the bundled CUDA 13 ptxas, now 13.4.59. Older targets keep the CUDA 12 ptxas ([#12035](https://github.com/triton-lang/triton/pull/12035), [#11669](https://github.com/triton-lang/triton/pull/11669))
- **Conversions:** FP16, BF16, FP8, and FP4 conversions use native instructions on Hopper and Blackwell. Fixed rounding in round-to-nearest-even conversions to `float8e5` and `float8e4nv` ([#11340](https://github.com/triton-lang/triton/pull/11340), [#11647](https://github.com/triton-lang/triton/

## retained text (truncated)



```
# Triton 3.9.0 Release Notes

## Table of Contents

- [Dialect & Frontend](#dialect--frontend-39)
- [Backend & Compiler](#backend--compiler-39)
- [AMD/HIP Backend](#amdhip-backend-39)
- [NVIDIA Backend](#nvidia-backend-39)
- [Gluon & Layout Improvements](#gluon--layout-improvements-39)
- [Kernels & Benchmarks](#kernels--benchmarks-39)
- [Proton Profiling](#proton-profiling-39)
- [Testing & CI](#testing--ci-39)
- [Build & Infrastructure](#build--infrastructure-39)
- [Documentation](#documentation-39)
- [Deprecations](#deprecations-39)
- [Breaking Changes](#breaking-changes-39)
- [Contributors](#contributors-39)

---

## Dialect & Frontend<a id="dialect--frontend-39"></a>

### New Features

- **Atomic loads and stores:** Added `tl.atomic_load` and `tl.atomic_store`, also available in Gluon and the interpreter. Loads default to `acquire` and stores default to `release` ([#11591](https://github.com/triton-lang/triton/pull/11591), [#11650](https://github.com/triton-lang/triton/pull/11650), [#11676](https://github.com/triton-lang/triton/pull/11676), [#11821](https://github.com/triton-lang/triton/pull/11821))
- **`tl.atomic_poll`:** Added `tl.atomic_poll`, which waits until the value at a pointer equals an expected value and returns `False` on timeout ([#10723](https://github.com/triton-lang/triton/pull/10723), [#10755](https://github.com/triton-lang/triton/pull/10755), [#11540](https://github.com/triton-lang/triton/pull/11540))
- **Approximate division:** Added `tl.fdiv(..., approx=True)` for fast FP32 division ([#11900](https://github.com/triton-lang/triton/pull/11900))
- **Language syntax:** Kernels support chained comparisons such as `0 <= x <= 8`, keyword-only parameters, tuple unpacking in list comprehensions, and operator methods on aggregate types ([#12002](https://github.com/triton-lang/triton/pull/12002), [#11418](https://github.com/triton-lang/triton/pull/11418), [#11076](https://github.com/triton-lang/triton/pull/11076), [#10990](https://github.com/triton-lang/triton/pull/10990))
- **Integer `tl.clamp`:** `tl.clamp` accepts integer tensors ([#12005](https://github.com/triton-lang/triton/pull/12005))

### Improvements

- **Earlier errors:** Mismatched tuple unpacking, invalid `tl.topk` `k` values, `tl.cat` shape mismatches, and non-32-bit `tl.histogram` inputs now fail at compile time ([#11053](https://github.com/triton-lang/triton/pull/11053), [#11074](https://github.com/triton-lang/triton/pull/11074), [#11353](https://github.com/triton-lang/triton/pull/11353), [#10891](https://github.com/triton-lang/triton/pull/10891))
- **Hints through helpers:** Alignment and contiguity hints on values returned from inlined helpers are now kept ([#11150](https://github.com/triton-lang/triton/pull/11150))
- **Compile time:** Deeply nested loops compile about 10x faster at nesting depth 8 ([#11369](https://github.com/triton-lang/triton/pull/11369))

### Bug Fixes

- **Wrong results:** Fixed `tl.softmax` with `dim` other than 0, integer stores to boolean pointers, list-comprehension `if` filters, negative-zero constants, and stale compilations when a captured constexpr shares a global's name ([#11409](https://github.com/triton-lang/triton/pull/11409), [#11853](https://github.com/triton-lang/triton/pull/11853), [#10888](https://github.com/triton-lang/triton/pull/10888), [#11426](https://github.com/triton-lang/triton/pull/11426), [#11224](https://github.com/triton-lang/triton/pull/11224))
- **E8M0 scale decoding:** E8M0 `0x00` now decodes to `2^-127` and `0xFF` to NaN (previously zero and infinity), which changes results for kernels that use those values ([#11624](https://github.com/triton-lang/triton/pull/11624))
- **Forked processes:** A forked child process no longer unloads GPU modules that belong to its parent ([#11128](https://github.com/triton-lang/triton/pull/11128))
- **Interpreter:** `TRITON_INTERPRET=1` now matches compiled kernels for integer overflow, `tl.fma`, `tl.umulhi`, and `tl.histogram` ([#10923](https://github.com/triton-lang/triton/pull/10923), [#11194](https://github.com/triton-lang/triton/pull/11194), [#11210](https://github.com/triton-lang/triton/pull/11210), [#11390](https://github.com/triton-lang/triton/pull/11390), [#11424](https://github.com/triton-lang/triton/pull/11424), [#11531](https://github.com/triton-lang/triton/pull/11531), [#11641](https://github.com/triton-lang/triton/pull/11641), [#11403](https://github.com/triton-lang/triton/pull/11403), [#11218](https://github.com/triton-lang/triton/pull/11218), [#11751](https://github.com/triton-lang/triton/pull/11751))

---

## Backend & Compiler<a id="backend--compiler-39"></a>

### LLVM Updates

- **LLVM pin:** Updated LLVM to llvm/llvm-project@b010a18d. AMD code generation now uses its own LLVM pin; see [AMD/HIP Backend](#amdhip-backend-39) ([#10739](https://github.com/triton-lang/triton/pull/10739), [#11163](https://github.com/triton-lang/triton/pull/11163), [#11485](https://github.com/triton-lang/triton/pull/11485))

### Multi-CTA / Multicast / TMA

- **Multi-CTA kernels:** Fixed missing cluster synchronization and proxy fences. Multi-CTA kernels now support local gather from subslices, `tl.histogram`, and ordered atomics in warp-specialized kernels ([#10656](https://github.com/triton-lang/triton/pull/10656), [#10735](https://github.com/triton-lang/triton/pull/10735), [#10843](https://github.com/triton-lang/triton/pull/10843), [#10975](https://github.com/triton-lang/triton/pull/10975), [#10559](https://github.com/triton-lang/triton/pull/10559), [#11775](https://github.com/triton-lang/triton/pull/11775), [#10914](https://github.com/triton-lang/triton/pull/10914))
- **Unaligned TMA:** Fixed wrong results when per-stage TMA buffers are not 128-byte aligned ([#10898](https://github.com/triton-lang/triton/pull/10898))

### Code Generation & Analysis

- **Atomic ordering:** Atomics with `acquire`, `release`, or `acq_rel` semantics now insert CTA-wide barriers so the ordering covers the whole block. Read-modify-write and CAS atomics default to `acq_rel`; pass `sem="relaxed"` to avoid the barriers ([#10816](https://github.com/triton-lang/triton/pull/10816), [#11022](https://github.com/triton-lang/triton/pull/11022))
- **Barrier insertion:** Barrier insertion was rewritten as a standalone pass that adds missing barriers and fences and removes redundant ones ([#11389](https://github.com/triton-lang/triton/pull/11389), [#11391](https://github.com/triton-lang/triton/pull/11391), [#11396](https://github.com/triton-lang/triton/pull/11396), [#11401](https://github.com/triton-lang/triton/pull/11401), [#11411](https://github.com/triton-lang/triton/pull/11411), [#11414](https://github.com/triton-lang/triton/pull/11414), [#11431](https://github.com/triton-lang/triton/pull/11431), [#11462](https://github.com/triton-lang/triton/pull/11462), [#11468](https://github.com/triton-lang/triton/pull/11468), [#11479](https://github.com/triton-lang/triton/pull/11479), [#10087](https://github.com/triton-lang/triton/pull/10087), [#11324](https://github.com/triton-lang/triton/pull/11324), [#11464](https://github.com/triton-lang/triton/pull/11464), [#11499](https://github.com/triton-lang/triton/pull/11499), [#11597](https://github.com/triton-lang/triton/pull/11597))
- **Batched matmul:** On NVIDIA, batched MMAv2 dots are up to 40x faster ([#11822](https://github.com/triton-lang/triton/pull/11822))
- **Miscompile fixes:** Fixed wrong results in nested-loop flattening, reductions, scans, `tl.histogram`, gathers, axis analysis, and optimizations that moved or duplicated loads ([#11521](https://github.com/triton-lang/triton/pull/11521), [#11688](https://github.com/triton-lang/triton/pull/11688), [#11617](https://github.com/triton-lang/triton/pull/11617), [#11621](https://github.com/triton-lang/triton/pull/11621), [#10744](https://github.com/triton-lang/triton/pull/10744), [#11186](https://github.com/triton-lang/triton/pull/11186), [#11269](https://github.com/triton-lang/triton/pull/11269), [#11682](https://github.com/triton-lang/triton/pull/11682), [#11842](https://github.com/triton-lang/triton/pull/11842), [#11637](https://github.com/triton-lang/triton/pull/11637), [#11981](https://github.com/triton-lang/triton/pull/11981), [#11990](https://github.com/triton-lang/triton/pull/11990), [#11943](https://github.com/triton-lang/triton/pull/11943), [#12011](https://github.com/triton-lang/triton/pull/12011), [#11767](https://github.com/triton-lang/triton/pull/11767), [#10838](https://github.com/triton-lang/triton/pull/10838), [#11546](https://github.com/triton-lang/triton/pull/11546), [#11548](https://github.com/triton-lang/triton/pull/11548), [#11337](https://github.com/triton-lang/triton/pull/11337), [#11314](https://github.com/triton-lang/triton/pull/11314), [#11906](https://github.com/triton-lang/triton/pull/11906), [#11975](https://github.com/triton-lang/triton/pull/11975), [#11240](https://github.com/triton-lang/triton/pull/11240))

### Sanitizers & Debugging

- **ConSan:** ConSan catches more races and reports fewer false positives. It compiles and runs about 2.5x faster on one reported workload ([#10668](https://github.com/triton-lang/triton/pull/10668), [#10736](https://github.com/triton-lang/triton/pull/10736), [#10784](https://github.com/triton-lang/triton/pull/10784), [#10864](https://github.com/triton-lang/triton/pull/10864), [#10948](https://github.com/triton-lang/triton/pull/10948), [#11045](https://github.com/triton-lang/triton/pull/11045), [#11010](https://github.com/triton-lang/triton/pull/11010), [#11162](https://github.com/triton-lang/triton/pull/11162), [#11216](https://github.com/triton-lang/triton/pull/11216), [#11244](https://github.com/triton-lang/triton/pull/11244), [#11310](https://github.com/triton-lang/triton/pull/11310), [#11200](https://github.com/triton-lang/triton/pull/11200), [#11202](https://github.com/triton-lang/triton/pull/11202), [#11175](https://github.com/triton-lang/triton/pull/11175), [#11268](https://github.com/triton-lang/triton/pull/11268), [#11436](https://github.com/triton-lang/triton/pull/11436), [#11459](https://github.com/triton-lang/triton/pull/11459), [#11626](https://github.com/triton-lang/triton/pull/11626), [#11852](https://github.com/triton-lang/triton/pull/11852), [#11857](https://github.com/triton-lang/triton/pull/11857), [#11138](https://github.com/triton-lang/triton/pull/11138), [#11446](https://github.com/triton-lang/triton/pull/11446), [#11447](https://github.com/triton-lang/triton/pull/11447), [#11448](https://github.com/triton-lang/triton/pull/11448), [#11449](https://github.com/triton-lang/triton/pull/11449), [#11450](https://github.com/triton-lang/triton/pull/11450), [#11451](https://github.com/triton-lang/triton/pull/11451), [#11452](https://github.com/triton-lang/triton/pull/11452), [#11642](https://github.com/triton-lang/triton/pull/11642))
- **GSan:** GSan supports programmatic dependent launch, CUDA graphs built with explicit dependencies, and multi-CTA kernels, and runs up to 3.8x faster on MoE ([#10677](https://github.com/triton-lang/triton/pull/10677), [#10944](https://github.com/triton-lang/triton/pull/10944), [#11020](https://github.com/triton-lang/triton/pull/11020), [#11135](https://github.com/triton-lang/triton/pull/11135), [#11199](https://github.com/triton-lang/triton/pull/11199), [#11303](https://github.com/triton-lang/triton/pull/11303), [#11172](https://github.com/triton-lang/triton/pull/11172), [#11187](https://github.com/triton-lang/triton/pull/11187), [#11297](https://github.com/triton-lang/triton/pull/11297), [#11351](https://github.com/triton-lang/triton/pull/11351), [#12023](https://github.com/triton-lang/triton/pull/12023))
- **FpSan:** FpSan supports `tl.clamp`, pure inline assembly, and Gluon embedding ops. Narrowing casts now fold discarded bits by default; set `TRITON_FPSAN_HOMOMORPHIC_CASTS=1` to keep truncation ([#10707](https://github.com/triton-lang/triton/pull/10707), [#10627](https:
```



document fetched 2026-10-28 02:33
