Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio int8 GEMM: keep sm8x GPUs on cuBLAS under Triton 3.8+, where every shipped tile spills (#17036)

Daniel Han711e29e9f59a
diff --git a/studio/backend/core/inference/diffusion_int8_gemm.py b/studio/backend/core/inference/diffusion_int8_gemm.py
index c67e5fa61e3..3e26248e44e 100644
--- a/studio/backend/core/inference/diffusion_int8_gemm.py
+++ b/studio/backend/core/inference/diffusion_int8_gemm.py
@@ -1,10 +1,10 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-"""int8 GEMM with torchao dequant epilogue in Triton, bf16 out; sm80/sm89/sm120 measured, other sm80+ behind a
-bit-exact probe and a latency check (UNSLOTH_DIFFUSION_INT8_GEMM_UNLISTED=0 keeps them stock), sm100 / sm103 stock
-(ROCm and CPU never reach it); Linear output bit-identical, compiled block not end to end;
-sm120 loads its operands through TMA descriptors (UNSLOTH_DIFFUSION_INT8_GEMM_TMA=0 keeps pointer loads);
+"""int8 GEMM with torchao dequant epilogue in Triton, bf16 out; sm80/sm89/sm120 measured (every sm8x stock on
+Triton >= 3.8, whose int8 dot spills there), other sm80+ behind a bit-exact probe and a latency check
+(UNSLOTH_DIFFUSION_INT8_GEMM_UNLISTED=0 keeps them stock), sm100 / sm103 stock (ROCm and CPU never reach it);
+Linear output bit-identical, compiled block not end to end; sm120 loads its operands through TMA descriptors (UNSLOTH_DIFFUSION_INT8_GEMM_TMA=0 keeps pointer loads);
UNSLOTH_DIFFUSION_INT8_GEMM=0 disables."""
from __future__ import annotations
@@ -137,6 +137,23 @@ def _triton_version_ok(version: Optional[str] = None) -> bool:
return bool(match) and (int(match.group(1)), int(match.group(2))) >= _MIN_TRITON
+@lru_cache(maxsize = None)
+def _triton_at_least(major: int, minor: int) -> bool:
+ try:
+ import triton
+ except Exception: # noqa: BLE001
+ return False
+ import re
+
+ match = re.match(r"(\d+)\.(\d+)", str(triton.__version__))
+ return bool(match) and (int(match.group(1)), int(match.group(2))) >= (major, minor)
+
+
+def _triton38_sm8x(cap: tuple) -> bool:
+ """Triton >= 3.8 spills sm8x int8 tiles (triton-lang/triton#12486); even spill-free ones trail cuBLAS."""
+ return cap[0] == 8 and _triton_at_least(3, 8)
+
+
def unlisted_mode() -> str:
"""'off', 'force' (no latency check) or 'auto' for the sm80+ archs without an ``_ARCH_CONFIG`` tile."""
gemm = int8_gemm_mode()
@@ -161,7 +178,10 @@ def unlisted_wanted(capability: Optional[tuple], mode: Optional[str] = None) ->
mode = unlisted_mode() if mode is None else mode
if mode == "off" or not unlisted_arch(capability):
return False
- return mode == "force" or (int(capability[0]), int(capability[1])) not in _UNLISTED_SKIP
+ if mode == "force":
+ return True
+ cap = (int(capability[0]), int(capability[1]))
+ return cap not in _UNLISTED_SKIP and not _triton38_sm8x(cap)
def arch_config(capability: Optional[tuple], mode: Optional[str] = None) -> Optional[tuple]:
@@ -171,6 +191,8 @@ def arch_config(capability: Optional[tuple], mode: Optional[str] = None) -> Opti
return None
cap = (int(capability[0]), int(capability[1]))
cfg = _ARCH_CONFIG.get(cap)
+ if mode == "auto" and _triton38_sm8x(cap):
+ return None
if cfg is None and mode == "force" and cap >= (8, 0):
return _FALLBACK_CONFIG
return cfg
diff --git a/studio/backend/tests/test_diffusion_int8_gemm.py b/studio/backend/tests/test_diffusion_int8_gemm.py
index ac33a406ab4..808f53cc82a 100644
--- a/studio/backend/tests/test_diffusion_int8_gemm.py
+++ b/studio/backend/tests/test_diffusion_int8_gemm.py
@@ -50,7 +50,8 @@ def _clean(monkeypatch):
((12, 0), True),
],
)
-def test_arch_gate_auto(cap, on):
+def test_arch_gate_auto(monkeypatch, cap, on):
+ monkeypatch.setattr(g8, "_triton_at_least", lambda major, minor: False) # the pre-3.8 table
assert (g8.arch_config(cap, "auto") is not None) is on
@@ -68,6 +69,52 @@ def test_force_enables_unmeasured_sm80_plus_only(monkeypatch):
assert g8.arch_config((7, 5)) is None
[email protected]("newer", [False, True])
+def test_triton38_keeps_every_sm8x_stock(monkeypatch, newer):
+ monkeypatch.setattr(g8, "_triton_at_least", lambda major, minor: newer)
+ for cap, cfg in g8._ARCH_CONFIG.items():
+ stock = newer and cap[0] == 8
+ assert g8.arch_config(cap, "auto") == (None if stock else cfg)
+ assert g8.arch_config(cap, "force") == cfg # forcing still runs the kernel
+ assert not g8.unlisted_arch(cap) and not g8.unlisted_wanted(cap, "auto")
+ for cap in ((8, 6), (8, 7)): # unlisted sm8x: no probe / latency check either
+ assert g8.unlisted_wanted(cap, "auto") is not newer
+ assert g8.unlisted_wanted(cap, "force")
+ for cap in ((9, 0), (12, 1)): # no spill there: unchanged
+ assert g8.unlisted_wanted(cap, "auto")
+ assert g8.arch_config((12, 0), "auto") == g8._ARCH_CONFIG[(12, 0)]
+ assert g8._gates((8, 9))[0] is not newer
+
+
+ "version, newer", [("3.7.1", False), ("3.8.0", True), ("3.9.0+git1", True), ("4.0", True)]
+)
+def test_triton_at_least_parses_the_installed_version(monkeypatch, version, newer):
+ import sys
+
+ monkeypatch.setitem(sys.modules, "triton", types.SimpleNamespace(__version__ = version))
+ g8._triton_at_least.cache_clear()
+ try:
+ assert g8._triton_at_least(3, 8) is newer
+ finally:
+ g8._triton_at_least.cache_clear()
+
+
+def test_triton38_runtime_turns_sm80_sm89_stock(monkeypatch):
+ import sys
+
+ monkeypatch.setitem(sys.modules, "triton", types.SimpleNamespace(__version__ = "3.8.0"))
+ clear = getattr(getattr(g8, "_triton_at_least", None), "cache_clear", lambda: None)
+ clear()
+ try:
+ assert g8.arch_config((8, 0), "auto") is None
+ assert g8.arch_config((8, 9), "auto") is None
+ assert not g8.unlisted_wanted((8, 6), "auto")
+ assert g8.arch_config((12, 0), "auto") == g8._ARCH_CONFIG[(12, 0)]
+ finally:
+ clear()
+
+
def test_install_is_a_noop_when_killed(monkeypatch):
monkeypatch.setenv(g8.INT8_GEMM_ENV, "0")
lin = torch.nn.Linear(64, 64)
diff --git a/studio/backend/tests/test_diffusion_int8_gemm_unlisted.py b/studio/backend/tests/test_diffusion_int8_gemm_unlisted.py
index 1bdf4a7b274..d6e0240be46 100644
--- a/studio/backend/tests/test_diffusion_int8_gemm_unlisted.py
+++ b/studio/backend/tests/test_diffusion_int8_gemm_unlisted.py
@@ -42,6 +42,8 @@ def _reset():
def _clean(monkeypatch, tmp_path):
for name in _ENVS:
monkeypatch.delenv(name, raising = False)
+ # pre-Triton-3.8 gate; 3.8+ is in test_diffusion_int8_gemm.py
+ monkeypatch.setattr(g8, "_triton_at_least", lambda major, minor: False)
monkeypatch.setattr(g8, "_race_cache_file", lambda: tmp_path / g8._RACE_CACHE_NAME)
monkeypatch.setattr(g8, "_race_key", lambda index: f"card{index}")
_reset()