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: extend the ROCm MIOpen cutoff and Qwen-Image-2.1 query chunking beyond gfx1151 (#12471)

oobabooga96c8bd176fbd
diff --git a/studio/backend/core/inference/diffusion_qwenimage21_rocm.py b/studio/backend/core/inference/diffusion_qwenimage21_rocm.py
index 14ac8a2659b..9a5f73b936d 100644
--- a/studio/backend/core/inference/diffusion_qwenimage21_rocm.py
+++ b/studio/backend/core/inference/diffusion_qwenimage21_rocm.py
@@ -1,11 +1,9 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-"""Bound large Qwen-Image-2.1 query batches on measured ROCm GPUs.
+"""Speed up large Qwen-Image-2.1 ROCm attention with query chunks.
-On gfx1151, native flash attention slows sharply above 8192 image tokens. Query
-rows are independent for unmasked, noncausal attention, so each chunk can retain
-all keys and values. Prefill, masks and explicit alternative backends stay stock.
+Unmasked, noncausal queries are independent; each chunk keeps all keys and values.
"""
from __future__ import annotations
@@ -17,6 +15,8 @@ from typing import Any
QUERY_CHUNK_ENV = "UNSLOTH_QWEN_IMAGE_ROCM_QUERY_CHUNKING"
QUERY_CHUNK_SIZE = 512
MIN_QUERY_TOKENS = 8192
+# gfx1151 and gfx1201 are measured; gfx1200 runs gfx1201's AOTriton kernels and tuning.
+CHUNKED_ARCHES = frozenset({"gfx1151", "gfx1200", "gfx1201"})
def _device_supported(target: Any) -> bool:
@@ -35,7 +35,9 @@ def _device_supported(target: Any) -> bool:
props = torch.cuda.get_device_properties(
torch.cuda.current_device() if ordinal is None else ordinal
)
- return str(getattr(props, "gcnArchName", "")).split(":", 1)[0] == "gfx1151"
+ from utils.hardware.hardware import _props_gfx_arch
+
+ return _props_gfx_arch(props) in CHUNKED_ARCHES
def _native_backend(processor: Any) -> bool:
diff --git a/studio/backend/tests/test_diffusion_qwenimage21_rocm.py b/studio/backend/tests/test_diffusion_qwenimage21_rocm.py
index d07bd0b91cd..fcde0145825 100644
--- a/studio/backend/tests/test_diffusion_qwenimage21_rocm.py
+++ b/studio/backend/tests/test_diffusion_qwenimage21_rocm.py
@@ -13,8 +13,12 @@ from core.inference import diffusion_qwenimage21_rocm as chunks
[
("gfx1151", "auto", True),
("gfx1151:sramecc-", "auto", True),
+ ("gfx1201", "auto", True),
+ ("gfx1200", "auto", True),
("gfx1100", "auto", False),
+ ("gfx1150", "auto", False),
("gfx942", "auto", False),
+ ("gfx1201", "0", False),
("gfx1100", "1", True),
("gfx1151", "0", False),
],
diff --git a/studio/backend/tests/test_rocm_miopen_cutoff.py b/studio/backend/tests/test_rocm_miopen_cutoff.py
index 539b3e30ff0..6aa6bec3157 100644
--- a/studio/backend/tests/test_rocm_miopen_cutoff.py
+++ b/studio/backend/tests/test_rocm_miopen_cutoff.py
@@ -19,8 +19,18 @@ def isolated_environment(monkeypatch):
[
(["gfx1151"], "1"),
(["gfx1151:xnack-", "gfx1151"], "1"),
- (["gfx1100"], None),
- (["gfx1151", "gfx1100"], None),
+ (["gfx1100"], "1"),
+ (["gfx1102"], "1"),
+ (["gfx1150"], "1"),
+ (["gfx1200"], "1"),
+ (["gfx1201"], "1"),
+ (["gfx1151", "gfx1100", "gfx1201"], "1"),
+ (["gfx1030"], None),
+ (["gfx90a"], None),
+ (["gfx942:sramecc+:xnack-"], None),
+ (["gfx950"], None),
+ (["gfx1250"], None),
+ (["gfx1201", "gfx942"], None),
(["gfx1151", ""], None),
([], None),
],
diff --git a/studio/backend/utils/desktop_shell_env.py b/studio/backend/utils/desktop_shell_env.py
index f956061cb7f..9c47d99bf6b 100644
--- a/studio/backend/utils/desktop_shell_env.py
+++ b/studio/backend/utils/desktop_shell_env.py
@@ -48,7 +48,7 @@ ROCM_SHELL_ENV_ALLOWLIST: tuple[str, ...] = (
"MIOPEN_USER_DB_PATH",
"MIOPEN_CUSTOM_CACHE_DIR",
"MIOPEN_FIND_MODE",
- # A shell opt-out (0) or opt-in must beat hardware.py's gfx1151 default.
+ # Preserve shell overrides, including an explicit opt-out (0).
"MIOPEN_SEARCH_CUTOFF",
"HIP_VISIBLE_DEVICES",
"ROCR_VISIBLE_DEVICES",
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index fd5c7f54965..6e775351bb9 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -1407,21 +1407,28 @@ def _xpu_device_name_or_placeholder(torch) -> str:
return "<unavailable>"
+# RDNA 3/3.5/4 only; gfx1250 is Instinct, so avoid a broad gfx12 prefix.
+_MIOPEN_SEARCH_CUTOFF_ARCH_PREFIXES = ("gfx110", "gfx115", "gfx120")
+
+
def _configure_rocm_miopen(torch) -> None:
if "MIOPEN_SEARCH_CUTOFF" in os.environ:
return
try:
count = torch.cuda.device_count()
- # Process-wide in MIOpen: any untested arch on a mixed host keeps the default.
- if not count or any(
- _props_gfx_arch(torch.cuda.get_device_properties(i)) != "gfx1151" for i in range(count)
+ # MIOpen's cutoff is process-wide, so every visible GPU must qualify.
+ if not count or not all(
+ _props_gfx_arch(torch.cuda.get_device_properties(i)).startswith(
+ _MIOPEN_SEARCH_CUTOFF_ARCH_PREFIXES
+ )
+ for i in range(count)
):
return
except Exception as exc:
logger.debug("MIOpen search cutoff device probe failed: %s", exc)
return
os.environ.setdefault("MIOPEN_SEARCH_CUTOFF", "1")
- logger.info("ROCm gfx1151: enabled MIOpen search cutoff (MIOPEN_SEARCH_CUTOFF=1)")
+ logger.info("ROCm RDNA: enabled MIOpen search cutoff (MIOPEN_SEARCH_CUTOFF=1)")
def _detect_hardware_locked() -> DeviceType: