more accurate rope fallback (#2792)

This commit is contained in:
Awni Hannun
2025-11-19 06:07:21 -08:00
committed by GitHub
parent 6f35017d1b
commit f46877bc08
2 changed files with 31 additions and 9 deletions

View File

@@ -416,23 +416,25 @@ array rope(
if (offset.size() > 1) { if (offset.size() > 1) {
offset = expand_dims(offset, {-1, -2}, s); offset = expand_dims(offset, {-1, -2}, s);
} }
auto positions = auto positions = multiply(
multiply(add(arange(x.shape(2), t, s), offset, s), array(scale, t), s); add(arange(x.shape(2), float32, s), offset, s),
array(scale, float32),
s);
auto default_inv_freqs = [&s, &t, base, half_dims]() { auto default_inv_freqs = [&s, base, half_dims]() {
return exp( return exp(
multiply( multiply(
arange(0, -half_dims, -1, t, s), arange(0, -half_dims, -1, float32, s),
array(std::log(base) / half_dims, t), array(std::log(base) / half_dims, float32),
s), s),
s); s);
}; };
auto inv_freqs = inputs.size() == 3 ? astype(reciprocal(inputs[2], s), t, s) auto inv_freqs =
: default_inv_freqs(); inputs.size() == 3 ? reciprocal(inputs[2], s) : default_inv_freqs();
auto theta = multiply(expand_dims(positions, -1, s), inv_freqs, s); auto theta = multiply(expand_dims(positions, -1, s), inv_freqs, s);
auto coss = cos(theta, s); auto coss = astype(cos(theta, s), t, s);
auto sins = sin(theta, s); auto sins = astype(sin(theta, s), t, s);
auto apply_rope = [forward, s]( auto apply_rope = [forward, s](
const array& x1, const array& x1,

View File

@@ -332,6 +332,26 @@ class TestFast(mlx_tests.MLXTestCase):
rx = rope_orig(x, dims, traditional, base, scale, offset) rx = rope_orig(x, dims, traditional, base, scale, offset)
self.assertLess(mx.abs(rx - rx_fast).max(), 1e-5) self.assertLess(mx.abs(rx - rx_fast).max(), 1e-5)
def test_rope_with_large_offset(self):
x = mx.random.normal(shape=(1, 1, 1024, 32))
rx_fp32 = mx.fast.rope(
x,
32,
traditional=False,
scale=1.0,
base=10000,
offset=4000,
)
rx_bf16 = mx.fast.rope(
x.astype(mx.bfloat16),
32,
traditional=False,
scale=1.0,
base=10000,
offset=4000,
)
self.assertLess((rx_fp32 - rx_bf16).abs().max(), 1e-1)
def test_rms_norm(self): def test_rms_norm(self):
# Per dtype absolute tolerance # Per dtype absolute tolerance
tolerances = {mx.float32: 1e-6, mx.float16: 1e-3, mx.bfloat16: 1e-2} tolerances = {mx.float32: 1e-6, mx.float16: 1e-3, mx.bfloat16: 1e-2}