diff options
author | Kawrakow <48489457+ikawrakow@users.noreply.github.com> | 2023-06-08 19:46:22 +0300 |
---|---|---|
committer | GitHub <noreply@github.com> | 2023-06-08 19:46:22 +0300 |
commit | 0f291e1f65c1d68201e71ce99c89562a36686b6d (patch) | |
tree | 5325c9bd1f8954db8862d7021331c8b60840b631 /ggml-metal.m | |
parent | 8fc8179919a11738910db07a800f2b176f8adf09 (diff) |
metal : Q6_K implementation (#1752)
* 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>
Diffstat (limited to 'ggml-metal.m')
-rw-r--r-- | ggml-metal.m | 17 |
1 files changed, 17 insertions, 0 deletions
diff --git a/ggml-metal.m b/ggml-metal.m index f2a637b..626ca87 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -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"); } |