zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-07-25 12:01:17 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

4 bit working

Browse Source
This commit is contained in:
Alex Barron 2024-10-22 19:20:45 -07:00
parent 5824626c0b
commit ef14b1e9c3
3 changed files with 24 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
benchmarks/python/sdpa_vector_bench.py
View File

@ -2,7 +2,7 @@ import mlx.core as mx
import numpy as np import numpy as np
from time_utils import time_fn from time_utils import time_fn
L = 30000 L = 16
H = 32 H = 32
H_k = 32 // 4 H_k = 32 // 4
D = 128 D = 128
@ -29,13 +29,15 @@ def sdpa(q, k, v):
v = mx.quantize(v) v = mx.quantize(v)
k = mx.dequantize(*k) k = mx.dequantize(*k)
v = mx.dequantize(*v) v = mx.dequantize(*v)
return mx.fast.scaled_dot_product_attention(q, k, v, scale=1.0) return mx.fast.scaled_dot_product_attention(q, k, v, scale=0.08, mask=None)
def quant_sdpa(q, k, v): def quant_sdpa(q, k, v):
k = mx.quantize(k) k = mx.quantize(k)
v = mx.quantize(v) v = mx.quantize(v)
return mx.fast.quantized_scaled_dot_product_attention(q, *k, *v, scale=1.0) return mx.fast.quantized_scaled_dot_product_attention(
q, *k, *v, scale=0.08, mask=None
)
def time_self_attention_primitives(q, k, v): def time_self_attention_primitives(q, k, v):
@ -52,9 +54,14 @@ def time_self_attention_quant_sdpa(q, k, v):
if __name__ == "__main__": if __name__ == "__main__":
mx.random.seed(3) mx.random.seed(3)
q = mx.random.uniform(shape=(1, H, 10, D)) # q = mx.random.uniform(shape=(1, H, 1, D))
k = mx.random.uniform(shape=(1, H_k, L, D)) # k = mx.random.uniform(shape=(1, H_k, L, D))
v = mx.random.uniform(shape=(1, H_k, L, D)) # v = mx.random.uniform(shape=(1, H_k, L, D))
q = mx.array(np.load("/Users/alexbarron/mlx-examples/llms/queries.npy"))
k = mx.array(np.load("/Users/alexbarron/mlx-examples/llms/keys.npy"))
v = mx.array(np.load("/Users/alexbarron/mlx-examples/llms/values.npy"))
print(q.dtype)
print(q.shape, k.shape, v.shape)
mx.eval(q, k, v) mx.eval(q, k, v)
k_quant = mx.quantize(k) k_quant = mx.quantize(k)
@ -66,12 +73,12 @@ if __name__ == "__main__":
# time_self_attention_primitives(q, k, v) # time_self_attention_primitives(q, k, v)
q_sdpa = quant_sdpa(q, k, v) q_sdpa = quant_sdpa(q, k, v)
print(q_sdpa) print(q_sdpa)
o_attention = attention(q, k, v) # o_attention = attention(q, k, v)
print(o_attention) # print(o_attention)
np.testing.assert_allclose(q_sdpa, o_attention, atol=1e-5) # np.testing.assert_allclose(q_sdpa, o_attention, atol=1e-5)
# o_sdpa = sdpa(q, k, v) o_sdpa = sdpa(q, k, v)
# print(o_sdpa) print(o_sdpa)
# np.testing.assert_allclose(q_sdpa, o_sdpa, atol=1e-5) np.testing.assert_allclose(q_sdpa, o_sdpa, atol=1e-5)
# print(o_sdpa[..., :64]) # print(o_sdpa[..., :64])
# print() # print()
# print(o_attention[..., :64]) # print(o_attention[..., :64])

5
mlx/backend/metal/kernels/sdpa_vector.h
View File

@ -178,9 +178,10 @@ template <typename T, int D, int group_size, int bits>
U shifts[4] = {1, 16, 256, 4096}; U shifts[4] = {1, 16, 256, 4096};
for (int i = 0; i < elem_per_thread; i++) { for (int i = 0; i < elem_per_thread; i++) {
// Shift by the appropriate amount here // Shift by the appropriate amount here
query_sum += queries[i];
U shift = shifts[i % 4]; U shift = shifts[i % 4];
q[i] = static_cast<U>(scale) * queries[i] / shift; q[i] = static_cast<U>(scale) * queries[i];
query_sum += q[i];
q[i] /= shift;
} }
for (int i = 0; i < elem_per_thread; i++) { for (int i = 0; i < elem_per_thread; i++) {
o[i] = 0; o[i] = 0;

4
mlx/fast.cpp
View File

@ -687,7 +687,6 @@ array quantized_scaled_dot_product_attention(
auto n_q_heads = queries.shape(-3); auto n_q_heads = queries.shape(-3);
auto n_kv_heads = keys.shape(-3); auto n_kv_heads = keys.shape(-3);
std::cout << "group bits " << group_size << " " << bits << std::endl;
auto out_shape = std::vector<int>( auto out_shape = std::vector<int>(
{queries.shape(0), queries.shape(1), queries.shape(2), out_dim}); {queries.shape(0), queries.shape(1), queries.shape(2), out_dim});
auto stream = to_stream(s); auto stream = to_stream(s);
@ -747,7 +746,8 @@ array quantized_scaled_dot_product_attention(
return std::vector<array>{out}; return std::vector<array>{out};
}; };
if (true) { int L = queries.shape(2);
if (L > 1) {
if (needs_mask) { if (needs_mask) {
return fallback( return fallback(
{queries, {queries,