ggml : Q4_2 ARM

[ggerganov]

ref #959

This is a reimplementation of #1026 by introducing new quantization type Q4_2

This PR implements only ARM NEON. The plan is to merge this soon and add rest of the SIMD implementations.
For now, no need for SIMD quantize / dequantize - will be added later when needed

[prusnak]

One nitpick. Debug contains:

llama_model_load_internal: ftype      = 5 (unknown, may not work)

Fix:

diff --git a/llama.cpp b/llama.cpp
index ef8ee20..dd970f7 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -838,6 +838,7 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_F16:  return "mostly F16";
         case LLAMA_FTYPE_MOSTLY_Q4_0: return "mostly Q4_0";
         case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1";
+        case LLAMA_FTYPE_MOSTLY_Q4_2: return "mostly Q4_2";
         case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
                                       return "mostly Q4_1, some F16";
         default:                      return "unknown, may not work";

[prusnak]

Benchmark on Macbook M1 16 GB:

7B q4_0: 75 ms/token
7B q4_2: 105 ms/token

[ggerganov]

Benchmark on Macbook M1 16 GB:

7B q4_0: 75 ms/token 7B q4_2: 105 ms/token

I guess this is with 4 threads?
I have been using only 8 threads and didn't realize the the slow down is bigger for less threads

[sw]

This should probably use ggml_is_quantized as well:
https://github.com/ggerganov/llama.cpp/blob/99092f2f21809fa4a2b68f0c16b0607410ed4bb1/ggml.c#L10720

Otherwise, looking great! I get 1s/token :-(

[prusnak]

I guess this is with 4 threads?

Yes, 8 threads are around 2-times slower on M1 (which has 4 performance cores and 4 efficiency cores).

More tests on M1 with the speed up in f30dbf9a8be4755f8d0bd9575ba5540ccf9335a9:

7B q4_0 4 threads: 75 ms/token
7B q4_0 8 threads: 135 ms/token
7B q4_1 4 threads: 122 ms/token
7B q4_1 8 threads: 240 ms/token
7B q4_2 4 threads: 89 ms/token
7B q4_2 8 threads: 180 ms/token

This is great!

[ggerganov]

I guess this is with 4 threads?

Yes, 8 threads are around 2-times slower on M1 (which has 4 performance cores and 4 efficiency cores).

More tests on M1 with the speed up in f30dbf9:

7B q4_0 4 threads: 75 ms/token 7B q4_0 8 threads: 135 ms/token 7B q4_1 4 threads: 122 ms/token 7B q4_1 8 threads: 240 ms/token 7B q4_2 4 threads: 89 ms/token 7B q4_2 8 threads: 180 ms/token

This is great!

Try again with ggml : optimize q4_2 using vmlaq_n_f32 + vmulq_n_f32

I just found that vmlaq_n_f32 can be used to speed up things significantly.
Also optimized the Q4_1 to be about ~56 ms / token on M1 Pro (branch: q4_1xq8_0) which is pretty good.
Only slightly slower compared to Q4_0 at ~48 ms / token.

I have a feeling that the next Q4_3 quantization (i.e. Q4_1 but with F16 factors) will be able to evaluate at ~60 ms / token and hopefully the perplexity will be very close to full F16 (i.e. ~6.00 for 7B). Which would be just perfect

[prusnak]

Post-merge tests (from master 77a7340):

I see no significant change from earlier tests for 7B q4_2 4 threads

7B q4_2 4 threads: 89 ms/token

But I see a small improvement for 7B q4_2 8 threads:

7B q4_2 8 threads: 180 -> 173 ms/token