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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""LFM3's remote Triton conv dies with ``PassManager::run failed`` where Triton cannot build for the card (gfx101x on Triton 3.9); use its torch path there."""
from __future__ import annotations
import sys
from typing import Any, Callable, Optional
from core.inference.rocm_bf16 import _device_gfx_arch, is_rocm_torch
_GATE = "can_use_lfm3_conv"
def _core_probe() -> Callable[[int], bool]:
try:
from unsloth.models._utils import triton_runs_on_device
except Exception: # noqa: BLE001 -- no probe: keep the remote code's own choice
return lambda index: True
return triton_runs_on_device
def _model_cuda_indices(model: Any) -> set[int]:
indices = set()
try:
for param in model.parameters():
device = param.device
if device.type == "cuda":
indices.add(device.index or 0)
except Exception: # noqa: BLE001 -- no parameters to read: nothing on a GPU to gate
pass
return indices
def _guarded(original: Any, runs_on: Callable[[int], bool]):
def gate(hidden_states, weight, *args, **kwargs):
# Module-wide gate: a model on a card Triton does support keeps the kernel.
device = getattr(hidden_states, "device", None)
if getattr(device, "type", None) != "cuda" or not runs_on(device.index or 0):
return False
return original(hidden_states, weight, *args, **kwargs)
gate._unsloth_triton_guard = True
return gate
def disable_lfm3_triton_conv_where_unsupported(
model: Any,
torch: Any,
log: Any = None,
runs_on: Optional[Callable[[int], bool]] = None,
) -> list[str]:
"""Make every loaded remote-code module's ``can_use_lfm3_conv`` refuse tensors on a ROCm card Triton cannot run on."""
if not is_rocm_torch(torch):
return []
runs_on = runs_on or _core_probe()
failing = sorted(index for index in _model_cuda_indices(model) if not runs_on(index))
if not failing:
return []
patched = []
for name, module in list(sys.modules.items()):
if not name.startswith("transformers_modules.") or module is None:
continue
original = getattr(module, _GATE, None)
if callable(original) and not getattr(original, "_unsloth_triton_guard", False):
setattr(module, _GATE, _guarded(original, runs_on))
patched.append(name)
if patched and log is not None:
archs = ", ".join(_device_gfx_arch(torch, index) or f"cuda:{index}" for index in failing)
log.info(
f"LFM3 Triton conv disabled on {archs} (Triton cannot build for it), "
f"using the torch conv: {', '.join(patched)}"
)
return patched