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.
"""The RoPE kernel's constexpr strides and head counts are fixed per model: a compiled forward and
backward over changing batch sizes and sequence lengths stays one graph and matches eager."""
import pytest
import torch
from real_accelerator import has_real_cuda # tests/_shared, on sys.path via tests/conftest.py
if not has_real_cuda():
pytest.skip(reason = "needs a CUDA device", allow_module_level = True)
pytest.importorskip("triton")
import unsloth # noqa: F401
from unsloth.kernels import rope_embedding as RE
# Batch never 1 or equal to a head count, seq never head_dim: torch's own 0/1 and duck-size
# specializations would recompile on any kernel.
N_HEADS, N_KV, HEAD_DIM = 8, 4, 64
def _qk(batch, seq):
torch.manual_seed(batch * 1000 + seq)
width = (N_HEADS + 2 * N_KV) * HEAD_DIM
qkv = torch.randn(batch, seq, width, device = "cuda", dtype = torch.float16, requires_grad = True)
Q = qkv[..., : N_HEADS * HEAD_DIM].view(batch, seq, N_HEADS, HEAD_DIM).transpose(1, 2)
K = qkv[..., N_HEADS * HEAD_DIM : (N_HEADS + N_KV) * HEAD_DIM]
K = K.view(batch, seq, N_KV, HEAD_DIM).transpose(1, 2)
return qkv, Q, K
@pytest.mark.gpu
@pytest.mark.parametrize("with_indices", [False, True])
def test_compiled_rope_keeps_one_graph_across_batch_and_seq(with_indices):
if not RE._TRACEABLE:
pytest.skip(reason = "RoPE ops are not traceable on this torch / Triton")
from torch._dynamo.utils import counters
freqs = torch.outer(
torch.arange(4096.0, device = "cuda"), torch.rand(HEAD_DIM // 2, device = "cuda")
)
emb = torch.cat((freqs, freqs), -1)
cos, sin = emb.cos().half(), emb.sin().half()
def rope(Q, K, idx):
q, k = RE.fast_rope_embedding(Q, K, cos, sin, idx)
return q.float().square().sum() + k.float().sum()
torch._dynamo.reset()
counters.clear()
compiled = torch.compile(rope, dynamic = True, fullgraph = True)
for batch, seq in [(3, 37), (5, 96), (6, 129), (7, 13), (3, 300)]:
idx = torch.randint(0, 4096, (batch, seq), device = "cuda") if with_indices else None
qkv, Q, K = _qk(batch, seq)
compiled(Q, K, idx).backward()
grad = qkv.grad
qkv, Q, K = _qk(batch, seq)
rope(Q, K, idx).backward()
torch.testing.assert_close(grad, qkv.grad, rtol = 1e-3, atol = 1e-3)
assert counters["stats"]["unique_graphs"] == 1
@pytest.mark.gpu
def test_compiled_rope_seq_major_grad_keeps_one_graph():
# dK as [b, h, d, s] (fp16 LoRA compiled block) must not specialize seq_len.
if not RE._TRACEABLE:
pytest.skip(reason = "RoPE ops are not traceable on this torch / Triton")
from torch._dynamo.utils import counters
freqs = torch.outer(
torch.arange(4096.0, device = "cuda"), torch.rand(HEAD_DIM // 2, device = "cuda")
)
emb = torch.cat((freqs, freqs), -1)
cos, sin = emb.cos().half(), emb.sin().half()
def rope(Q, K):
q, k = RE.fast_rope_embedding(Q, K, cos, sin, None)
return q.float().square().sum() + k.transpose(2, 3).contiguous().float().square().sum()
torch._dynamo.reset()
counters.clear()
compiled = torch.compile(rope, dynamic = True, fullgraph = True)
for batch, seq in [(3, 37), (5, 96), (6, 129), (7, 13), (3, 300)]:
qkv, Q, K = _qk(batch, seq)
compiled(Q, K).backward()
grad = qkv.grad
qkv, Q, K = _qk(batch, seq)
rope(Q, K).backward()
torch.testing.assert_close(grad, qkv.grad, rtol = 1e-2, atol = 1e-2)
assert counters["stats"]["unique_graphs"] == 1