vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

metal : Q6_K implementation (#1752)

Browse source 
* Metal implementation for Q4_K

Very slow for now:
42 ms / token, Q4_0 runs in 28 ms/token on my
30-core M2 Max GPU.

* Optimizing Q4_K on metal

The first token always takes longer, I guess because
the metal kernel is being jit-compiled.
So, using n = 128 to measure time.

At this point Q4_K takes 29.5 ms / token
compared to 27.2 ms / token for Q4_0.
Quite a bit better than the initial attempt,
but still not good enough.

* Optimizing q4_K metal dot some more

For n = 256 it is now 28.1 ms/token compared to
27 ms/token for q4_0.

* Fix after merge with master

* Metal implementation for Q6_K

Similar to the CUDA implementation.
No idea if this is the optimum for Metal, but the few
alternative variants I tried all had a lower performance.

We get 36.5 ms / token on M2 Max with 30 GPU cores.
This corresponds to ~200 GB/second throughput.

* clang-tidy : add config back

* Much better Q6_K implementation for metal

28.3 ms / token for 7B. Subtracting ~9 ms that is spent in
other compute graph operations, we are left with ~19 ms
for the matrix multiplications. The model is ~5.5 GB,
so we are getting 1000 / 19 * 5.5 = 290 GB/s!

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow 2023-06-08 19:46:22 +03:00 • committed by GitHub
parent 8fc8179919
commit 0f291e1f65
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 187 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

17
ggml-metal.m
View file

@ -50,10 +50,12 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(get_rows_f16);
GGML_METAL_DECL_KERNEL(get_rows_q4_0);
GGML_METAL_DECL_KERNEL(get_rows_q4_k);
GGML_METAL_DECL_KERNEL(get_rows_q6_k);
GGML_METAL_DECL_KERNEL(rms_norm);
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_k_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q6_k_f32);
GGML_METAL_DECL_KERNEL(rope);
GGML_METAL_DECL_KERNEL(cpy_f32_f16);
GGML_METAL_DECL_KERNEL(cpy_f32_f32);
@ -136,10 +138,12 @@ struct ggml_metal_context * ggml_metal_init(void) {
GGML_METAL_ADD_KERNEL(get_rows_f16);
GGML_METAL_ADD_KERNEL(get_rows_q4_0);
GGML_METAL_ADD_KERNEL(get_rows_q4_k);
GGML_METAL_ADD_KERNEL(get_rows_q6_k);
GGML_METAL_ADD_KERNEL(rms_norm);
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_k_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q6_k_f32);
GGML_METAL_ADD_KERNEL(rope);
GGML_METAL_ADD_KERNEL(cpy_f32_f16);
GGML_METAL_ADD_KERNEL(cpy_f32_f32);
@ -530,6 +534,15 @@ void ggml_metal_graph_compute(
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
} break;
case GGML_TYPE_Q6_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
} break;
default:
{
fprintf(stderr, "Asserting on type %d\n",(int)src0t);
@ -560,6 +573,9 @@ void ggml_metal_graph_compute(
} else if (src0t == GGML_TYPE_Q4_K) {
[encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else if (src0t == GGML_TYPE_Q6_K) {
[encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@ -576,6 +592,7 @@ void ggml_metal_graph_compute(
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
default: GGML_ASSERT(false && "not implemented");
}