llamafile : improve moe prompt eval speed on cpu #6840

jart · 2024-04-23T05:41:24Z

This change introduces a llamafile_mixmul() API that allows tinyBLAS to speed up "Mixture of Expert" models. On my Threadripper, Mixtral's 8x7b F16 weights now process prompts 2x faster. I'm also seeing a 60 percent improvement with Mixtral 8x22b Q4_0. The same applies to Q8_0, which is also supported by tinyBLAS. MoE models spend the majority of their time inside MUL_MAT_ID rather than MUL_MAT, which is why llamafile_sgemm was not able to help them before. llamafile_mixmul works by decomposing the mixmul operation into approximatively two sgemm calls.

hiepxanh · 2024-04-23T12:50:10Z

nice to see this PR <3 Thank you so much

USBhost · 2024-04-23T16:08:17Z

Does it also help the other K quants?

jart · 2024-04-23T17:31:35Z

@USBhost Unfortunately no. The K quants were designed to exploit under-utilization of CPU resources when doing matvecs. I tried copying and pasting the Q5_K_M code into a tinyBLAS 2-d block-tiling kernel, but the compiler wasn't able to unroll it it in a way that offered performance gains through instruction level parallelism. I've only been able to make the simpler quants work. It's a shame because I really like Q5_K_M, so it'd be great to see Iwan Kawrakow develop a new quant specifically for block-tiling.

USBhost · 2024-04-23T17:45:18Z

@USBhost Unfortunately no. The K quants were designed to exploit under-utilization of CPU resources when doing matvecs. I tried copying and pasting the Q5_K_M code into a tinyBLAS 2-d block-tiling kernel, but the compiler wasn't able to unroll it it in a way that offered performance gains through instruction level parallelism. I've only been able to make the simpler quants work. It's a shame because I really like Q5_K_M, so it'd be great to see Iwan Kawrakow develop a new quant specifically for block-tiling.

I see thanks for the explanation.

Side note: I would love a doc that explains the speed between flat 4_0 vs 4_1 vs K quants. Because I keep seeing the simple ones getting buffs.

jart · 2024-04-23T18:20:01Z

The tinyBLAS code upstreamed by Mozilla's llamafile project makes prompt processing go very fast for F32, F16, Q4_0, and Q8_0.

model	size	params	backend	threads	test	t/s
llama 1B F16	2.05 GiB	1.10 B	CPU	96	pp 512	2048.86 ± 6.52
llama 1B F16	2.05 GiB	1.10 B	CPU	96	tg 4	52.01 ± 0.06
llama 1B all F32	4.10 GiB	1.10 B	CPU	96	pp 512	1946.19 ± 21.26
llama 1B all F32	4.10 GiB	1.10 B	CPU	96	tg 4	39.75 ± 0.15
llama 1B Q2_K - Medium	411.41 MiB	1.10 B	CPU	96	pp 512	1273.08 ± 18.57
llama 1B Q2_K - Medium	411.41 MiB	1.10 B	CPU	96	tg 4	69.55 ± 0.22
llama 1B Q3_K - Large	563.42 MiB	1.10 B	CPU	96	pp 512	1109.08 ± 7.66
llama 1B Q3_K - Large	563.42 MiB	1.10 B	CPU	96	tg 4	66.83 ± 0.41
llama 1B Q3_K - Medium	522.30 MiB	1.10 B	CPU	96	pp 512	1161.17 ± 7.58
llama 1B Q3_K - Medium	522.30 MiB	1.10 B	CPU	96	tg 4	67.49 ± 0.10
llama 1B Q3_K - Small	475.51 MiB	1.10 B	CPU	96	pp 512	1052.25 ± 161.64
llama 1B Q3_K - Small	475.51 MiB	1.10 B	CPU	96	tg 4	68.30 ± 0.19
llama 1B Q4_0	606.53 MiB	1.10 B	CPU	96	pp 512	1418.41 ± 10.54
llama 1B Q4_0	606.53 MiB	1.10 B	CPU	96	tg 4	65.78 ± 0.24
llama 1B Q4_1	668.18 MiB	1.10 B	CPU	96	pp 512	884.68 ± 3.74
llama 1B Q4_1	668.18 MiB	1.10 B	CPU	96	tg 4	64.62 ± 0.05
llama 1B Q4_K - Medium	636.18 MiB	1.10 B	CPU	96	pp 512	1197.76 ± 11.42
llama 1B Q4_K - Medium	636.18 MiB	1.10 B	CPU	96	tg 4	66.04 ± 0.34
llama 1B Q4_K - Small	609.53 MiB	1.10 B	CPU	96	pp 512	1200.22 ± 10.06
llama 1B Q4_K - Small	609.53 MiB	1.10 B	CPU	96	tg 4	66.38 ± 0.27
llama 1B Q5_0	729.84 MiB	1.10 B	CPU	96	pp 512	1058.68 ± 10.52
llama 1B Q5_0	729.84 MiB	1.10 B	CPU	96	tg 4	63.10 ± 0.36
llama 1B Q5_1	791.50 MiB	1.10 B	CPU	96	pp 512	718.18 ± 127.77
llama 1B Q5_1	791.50 MiB	1.10 B	CPU	96	tg 4	62.07 ± 0.67
llama 1B Q5_K - Medium	745.11 MiB	1.10 B	CPU	96	pp 512	1055.78 ± 5.80
llama 1B Q5_K - Medium	745.11 MiB	1.10 B	CPU	96	tg 4	64.01 ± 0.16
llama 1B Q5_K - Small	729.84 MiB	1.10 B	CPU	96	pp 512	1048.20 ± 3.90
llama 1B Q5_K - Small	729.84 MiB	1.10 B	CPU	96	tg 4	64.32 ± 0.27
llama 1B Q6_K	860.86 MiB	1.10 B	CPU	96	pp 512	995.96 ± 183.61
llama 1B Q6_K	860.86 MiB	1.10 B	CPU	96	tg 4	62.67 ± 0.22
llama 1B Q8_0	1.09 GiB	1.10 B	CPU	96	pp 512	1430.38 ± 9.86
llama 1B Q8_0	1.09 GiB	1.10 B	CPU	96	tg 4	59.71 ± 0.14

Measured on AMD Ryzen Threadripper PRO 7995WX with TinyLlama 1.1B. This PR ensures those performance wins will happen for MoE models too.

jart · 2024-04-24T01:07:53Z

Note: I'm still in the process of testing this change and verifying it's correct on all compilers and architectures.

jart · 2024-04-25T01:02:31Z

OK I've worked out the remaining kinks. This code was just shipped as part of the llamafile 0.8 release. Thanks to this change, I'm seeing a 2x prompt eval speed increase across the board. My Threadripper now runs Mixtral 2x faster. My M2 Ultra runs Mixtral 2x faster on CPU. This change even pumps up the Raspberry Pi 5 to 78 tok/sec performance on non-MoE F16 models in case you want to buy a bag full of the things to build your next supercomputer. PTAL.

ikawrakow · 2024-04-26T14:25:55Z

It's a shame because I really like Q5_K_M, so it'd be great to see Iwan Kawrakow develop a new quant specifically for block-tiling.

@jart

I became intrigued by your assumption that block-tiling is required to speed up prompt processing for k-quants, so spent some time optimizing k-quant CPU matrix multiplications. I'm running on a 16-core Ryzen-7950X CPU, so have done just a better AVX2 implementation. Baseline for this CPU (using your PR) for a 7B LLaMA is

model	size	params	backend	threads	test	t/s
llama 7B Q4_0	3.56 GiB	6.74 B	CPU	16	pp 512	119.75 ± 0.38
llama 7B Q4_1	3.95 GiB	6.74 B	CPU	16	pp 512	63.80 ± 0.22
llama 7B Q5_0	4.33 GiB	6.74 B	CPU	16	pp 512	59.50 ± 0.11
llama 7B Q5_1	4.72 GiB	6.74 B	CPU	16	pp 512	56.28 ± 0.08

Q4_0 is much faster than the other legacy quants thanks to your tinyBLAS.

Here is what I get for k-quants

model	size	params	test	t/s (master)	t/s (optimized)	Speedup
llama 7B Q2_K - Small	2.16 GiB	6.74 B	pp 512	116.35 ± 0.08	162.66 ± 1.21	1.398 ± 0.009
llama 7B Q2_K - Medium	2.36 GiB	6.74 B	pp 512	100.49 ± 0.15	149.31 ± 0.48	1.486 ± 0.006
llama 7B Q3_K - Small	2.75 GiB	6.74 B	pp 512	82.01 ± 0.11	132.68 ± 0.21	1.618 ± 0.003
llama 7B Q3_K - Medium	3.07 GiB	6.74 B	pp 512	89.16 ± 0.10	136.53 ± 0.19	1.531 ± 0.003
llama 7B Q4_K - Small	3.59 GiB	6.74 B	pp 512	104.45 ± 0.15	144.87 ± 0.31	1.387 ± 0.003
llama 7B Q4_K - Medium	3.80 GiB	6.74 B	pp 512	101.50 ± 0.28	145.77 ± 0.33	1.436 ± 0.005
llama 7B Q5_K - Small	4.33 GiB	6.74 B	pp 512	73.72 ± 0.29	124.05 ± 0.14	1.682 ± 0.007
llama 7B Q5_K - Medium	4.45 GiB	6.74 B	pp 512	74.84 ± 0.13	126.52 ± 0.14	1.691 ± 0.004
llama 7B Q6_K	5.15 GiB	6.74 B	pp 512	81.38 ± 0.09	146.30 ± 0.15	1.798 ± 0.003

You favorite Q5_K_M quants are faster than Q4_0 with tinyBLAS with these changes :-)

There are 3 ingredients involved in this speedup:

Simdify Q8_K quantization. Quantization of activations is single-threaded in ggml. Quantization to Q8_0, needed by legacy quants, is already simdified, but quantization to Q8_K, required by k-quants, is not. This does not matter for token generation, but we get a 5-6% speedup for prompt processing.
Tweak some more the k-quant dot product kernels to reduce/eliminate dependencies between computational steps. I guess this is what you call "instruction level parallelism"
Last, but certainly not least, make use of the ability to do 2x2 matrix multiplications (rather than just dot products) that is already available in ggml (2 weight rows times two activation columns). This gives 20-40% speedup, depending on how costly it is to setup the bits as needed for the multiplication with the Q8_K quants.

We see Q4_K and Q5_K being ~2.2 times faster than their respective legacy counterparts Q4_1 and Q5_1.

jart · 2024-04-26T16:52:14Z

That's outstanding news @ikawrakow! I can't wait to see your code. Now I won't need to recommend the legacy quantization formats. Am I correct in understanding you used .nrows? Have you tried copying your optimized code into a tinyBLAS kernel? If you do that, then K quants might be able to surpass F16 and BF16 at evaluation.

ikawrakow · 2024-04-26T17:16:54Z

@jart

Yes, I used .nrows. I did that because it was not obvious to me how I can plug that into tinyBLAS. I could make a PR to your repository so you can do the integration into tinyBLAS?

jart · 2024-04-26T17:46:56Z

@ikawrakow Receiving a PR from you would honor the llamafile project. What you'd want to do is create a copy of tinyBLAS_Q0_ARM named like tinyBLAS_K5_ARM and then have your vec_dot code replace these specific lines. You'd then tune its mnpack() method to be smaller until eventually there's no stack spillage.

github-actions · 2024-04-26T18:15:48Z

📈 llama.cpp server for bench-server-baseline on Standard_NC4as_T4_v3 for phi-2-q4_0: 570 iterations 🚀

Expand details for performance related PR only

Concurrent users: 8, duration: 10m
HTTP request : avg=8180.47ms p(95)=19678.15ms fails=, finish reason: stop=511 truncated=59
Prompt processing (pp): avg=93.49tk/s p(95)=393.96tk/s
Token generation (tg): avg=34.01tk/s p(95)=49.86tk/s
ggml-org/models/phi-2/ggml-model-q4_0.gguf parallel=8 ctx-size=16384 ngl=33 batch-size=2048 ubatch-size=256 pp=1024 pp+tg=2048 branch=moe commit=bb3a5274c7c1efd883f7e57edb849c0394d2c91d

More

---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 570 iterations"
    y-axis "llamacpp:prompt_tokens_seconds"
    x-axis "llamacpp:prompt_tokens_seconds" 1715528716 --> 1715529342
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 351.7, 351.7, 351.7, 351.7, 351.7, 689.37, 689.37, 689.37, 689.37, 689.37, 566.85, 566.85, 566.85, 566.85, 566.85, 597.08, 597.08, 597.08, 597.08, 597.08, 658.88, 658.88, 658.88, 658.88, 658.88, 676.83, 676.83, 676.83, 676.83, 676.83, 680.7, 680.7, 680.7, 680.7, 680.7, 701.39, 701.39, 701.39, 701.39, 701.39, 722.37, 722.37, 722.37, 722.37, 722.37, 737.99, 737.99, 737.99, 737.99, 737.99, 769.43, 769.43, 769.43, 769.43, 769.43, 781.02, 781.02, 781.02, 781.02, 781.02, 781.84, 781.84, 781.84, 781.84, 781.84, 810.53, 810.53, 810.53, 810.53, 810.53, 790.89, 790.89, 790.89, 790.89, 790.89, 794.15, 794.15, 794.15, 794.15, 794.15, 792.32, 792.32, 792.32, 792.32, 792.32, 817.94, 817.94, 817.94, 817.94, 817.94, 817.43, 817.43, 817.43, 817.43, 817.43, 823.84, 823.84, 823.84, 823.84, 823.84, 828.15, 828.15, 828.15, 828.15, 828.15, 829.94, 829.94, 829.94, 829.94, 829.94, 824.55, 824.55, 824.55, 824.55, 824.55, 828.53, 828.53, 828.53, 828.53, 828.53, 840.82, 840.82, 840.82, 840.82, 840.82, 838.33, 838.33, 838.33, 838.33, 838.33, 837.49, 837.49, 837.49, 837.49, 837.49, 838.1, 838.1, 838.1, 838.1, 838.1, 843.67, 843.67, 843.67, 843.67, 843.67, 843.03, 843.03, 843.03, 843.03, 843.03, 844.97, 844.97, 844.97, 844.97, 844.97, 844.92, 844.92, 844.92, 844.92, 844.92, 854.04, 854.04, 854.04, 854.04, 854.04, 851.33, 851.33, 851.33, 851.33, 851.33, 838.0, 838.0, 838.0, 838.0, 838.0, 836.58, 836.58, 836.58, 836.58, 836.58, 839.07, 839.07, 839.07, 839.07, 839.07, 839.93, 839.93, 839.93, 839.93, 839.93, 840.09, 840.09, 840.09, 840.09, 840.09, 843.14, 843.14, 843.14, 843.14, 843.14, 850.48, 850.48, 850.48, 850.48, 850.48, 849.68, 849.68, 849.68, 849.68, 849.68, 847.76, 847.76, 847.76, 847.76, 847.76, 844.44, 844.44, 844.44, 844.44, 844.44, 849.1, 849.1, 849.1, 849.1, 849.1, 850.97, 850.97, 850.97, 850.97, 850.97, 852.84, 852.84, 852.84, 852.84, 852.84, 851.83, 851.83, 851.83, 851.83, 851.83, 856.22, 856.22, 856.22, 856.22, 856.22, 859.22, 859.22, 859.22, 859.22, 859.22, 858.41, 858.41, 858.41, 858.41, 858.41, 863.34, 863.34, 863.34, 863.34, 863.34, 865.05, 865.05, 865.05, 865.05, 865.05, 864.38, 864.38, 864.38, 864.38, 864.38, 865.17, 865.17, 865.17, 865.17, 865.17, 862.92, 862.92, 862.92, 862.92, 862.92, 862.6, 862.6, 862.6, 862.6, 862.6, 865.29, 865.29, 865.29, 865.29, 865.29, 865.53, 865.53, 865.53, 865.53, 865.53, 864.4, 864.4, 864.4, 864.4, 864.4, 863.31, 863.31, 863.31, 863.31]

More

---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 570 iterations"
    y-axis "llamacpp:predicted_tokens_seconds"
    x-axis "llamacpp:predicted_tokens_seconds" 1715528716 --> 1715529342
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 40.27, 40.27, 40.27, 40.27, 40.27, 41.21, 41.21, 41.21, 41.21, 41.21, 30.23, 30.23, 30.23, 30.23, 30.23, 30.84, 30.84, 30.84, 30.84, 30.84, 31.6, 31.6, 31.6, 31.6, 31.6, 31.93, 31.93, 31.93, 31.93, 31.93, 33.17, 33.17, 33.17, 33.17, 33.17, 34.04, 34.04, 34.04, 34.04, 34.04, 34.36, 34.36, 34.36, 34.36, 34.36, 34.46, 34.46, 34.46, 34.46, 34.46, 34.07, 34.07, 34.07, 34.07, 34.07, 34.14, 34.14, 34.14, 34.14, 34.14, 33.6, 33.6, 33.6, 33.6, 33.6, 33.21, 33.21, 33.21, 33.21, 33.21, 32.34, 32.34, 32.34, 32.34, 32.34, 32.15, 32.15, 32.15, 32.15, 32.15, 32.7, 32.7, 32.7, 32.7, 32.7, 32.48, 32.48, 32.48, 32.48, 32.48, 32.51, 32.51, 32.51, 32.51, 32.51, 32.48, 32.48, 32.48, 32.48, 32.48, 32.44, 32.44, 32.44, 32.44, 32.44, 32.73, 32.73, 32.73, 32.73, 32.73, 32.66, 32.66, 32.66, 32.66, 32.66, 32.88, 32.88, 32.88, 32.88, 32.88, 32.97, 32.97, 32.97, 32.97, 32.97, 32.83, 32.83, 32.83, 32.83, 32.83, 32.55, 32.55, 32.55, 32.55, 32.55, 32.65, 32.65, 32.65, 32.65, 32.65, 32.92, 32.92, 32.92, 32.92, 32.92, 32.92, 32.92, 32.92, 32.92, 32.92, 33.17, 33.17, 33.17, 33.17, 33.17, 33.27, 33.27, 33.27, 33.27, 33.27, 33.06, 33.06, 33.06, 33.06, 33.06, 32.84, 32.84, 32.84, 32.84, 32.84, 32.55, 32.55, 32.55, 32.55, 32.55, 32.51, 32.51, 32.51, 32.51, 32.51, 32.66, 32.66, 32.66, 32.66, 32.66, 32.77, 32.77, 32.77, 32.77, 32.77, 32.9, 32.9, 32.9, 32.9, 32.9, 32.93, 32.93, 32.93, 32.93, 32.93, 32.79, 32.79, 32.79, 32.79, 32.79, 31.75, 31.75, 31.75, 31.75, 31.75, 31.61, 31.61, 31.61, 31.61, 31.61, 30.39, 30.39, 30.39, 30.39, 30.39, 30.33, 30.33, 30.33, 30.33, 30.33, 30.43, 30.43, 30.43, 30.43, 30.43, 30.58, 30.58, 30.58, 30.58, 30.58, 30.72, 30.72, 30.72, 30.72, 30.72, 30.74, 30.74, 30.74, 30.74, 30.74, 30.73, 30.73, 30.73, 30.73, 30.73, 30.5, 30.5, 30.5, 30.5, 30.5, 30.4, 30.4, 30.4, 30.4, 30.4, 30.45, 30.45, 30.45, 30.45, 30.45, 30.63, 30.63, 30.63, 30.63, 30.63, 30.69, 30.69, 30.69, 30.69, 30.69, 30.83, 30.83, 30.83, 30.83, 30.83, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.92, 30.96, 30.96, 30.96, 30.96, 30.96, 31.05, 31.05, 31.05, 31.05]

Details

More

---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 570 iterations"
    y-axis "llamacpp:kv_cache_usage_ratio"
    x-axis "llamacpp:kv_cache_usage_ratio" 1715528716 --> 1715529342
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.12, 0.12, 0.12, 0.12, 0.12, 0.39, 0.39, 0.39, 0.39, 0.39, 0.18, 0.18, 0.18, 0.18, 0.18, 0.12, 0.12, 0.12, 0.12, 0.12, 0.2, 0.2, 0.2, 0.2, 0.2, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.11, 0.11, 0.11, 0.11, 0.11, 0.17, 0.17, 0.17, 0.17, 0.17, 0.19, 0.19, 0.19, 0.19, 0.19, 0.23, 0.23, 0.23, 0.23, 0.23, 0.18, 0.18, 0.18, 0.18, 0.18, 0.14, 0.14, 0.14, 0.14, 0.14, 0.34, 0.34, 0.34, 0.34, 0.34, 0.19, 0.19, 0.19, 0.19, 0.19, 0.13, 0.13, 0.13, 0.13, 0.13, 0.15, 0.15, 0.15, 0.15, 0.15, 0.19, 0.19, 0.19, 0.19, 0.19, 0.2, 0.2, 0.2, 0.2, 0.2, 0.18, 0.18, 0.18, 0.18, 0.18, 0.14, 0.14, 0.14, 0.14, 0.14, 0.31, 0.31, 0.31, 0.31, 0.31, 0.1, 0.1, 0.1, 0.1, 0.1, 0.16, 0.16, 0.16, 0.16, 0.16, 0.31, 0.31, 0.31, 0.31, 0.31, 0.22, 0.22, 0.22, 0.22, 0.22, 0.2, 0.2, 0.2, 0.2, 0.2, 0.1, 0.1, 0.1, 0.1, 0.1, 0.12, 0.12, 0.12, 0.12, 0.12, 0.11, 0.11, 0.11, 0.11, 0.11, 0.17, 0.17, 0.17, 0.17, 0.17, 0.21, 0.21, 0.21, 0.21, 0.21, 0.17, 0.17, 0.17, 0.17, 0.17, 0.29, 0.29, 0.29, 0.29, 0.29, 0.12, 0.12, 0.12, 0.12, 0.12, 0.16, 0.16, 0.16, 0.16, 0.16, 0.12, 0.12, 0.12, 0.12, 0.12, 0.08, 0.08, 0.08, 0.08, 0.08, 0.16, 0.16, 0.16, 0.16, 0.16, 0.37, 0.37, 0.37, 0.37, 0.37, 0.48, 0.48, 0.48, 0.48, 0.48, 0.5, 0.5, 0.5, 0.5, 0.5, 0.53, 0.53, 0.53, 0.53, 0.53, 0.19, 0.19, 0.19, 0.19, 0.19, 0.14, 0.14, 0.14, 0.14, 0.14, 0.11, 0.11, 0.11, 0.11, 0.11, 0.19, 0.19, 0.19, 0.19, 0.19, 0.1, 0.1, 0.1, 0.1, 0.1, 0.18, 0.18, 0.18, 0.18, 0.18, 0.25, 0.25, 0.25, 0.25, 0.25, 0.19, 0.19, 0.19, 0.19, 0.19, 0.12, 0.12, 0.12, 0.12, 0.12, 0.11, 0.11, 0.11, 0.11, 0.11, 0.08, 0.08, 0.08, 0.08, 0.08, 0.13, 0.13, 0.13, 0.13, 0.13, 0.12, 0.12, 0.12, 0.12, 0.12, 0.12, 0.12, 0.12, 0.12, 0.12, 0.14, 0.14, 0.14, 0.14, 0.14, 0.19, 0.19, 0.19, 0.19, 0.19, 0.15, 0.15, 0.15, 0.15, 0.15, 0.1, 0.1, 0.1, 0.1]

More

---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 570 iterations"
    y-axis "llamacpp:requests_processing"
    x-axis "llamacpp:requests_processing" 1715528716 --> 1715529342
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 6.0, 6.0, 6.0, 6.0, 6.0, 5.0, 5.0, 5.0, 5.0, 5.0, 6.0, 6.0, 6.0, 6.0, 6.0, 5.0, 5.0, 5.0, 5.0, 5.0, 4.0, 4.0, 4.0, 4.0, 4.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 1.0, 1.0, 1.0, 1.0, 1.0, 7.0, 7.0, 7.0, 7.0, 7.0, 5.0, 5.0, 5.0, 5.0, 5.0, 3.0, 3.0, 3.0, 3.0, 3.0, 7.0, 7.0, 7.0, 7.0, 7.0, 0.0, 0.0, 0.0, 0.0, 0.0, 6.0, 6.0, 6.0, 6.0, 6.0, 4.0, 4.0, 4.0, 4.0, 4.0, 6.0, 6.0, 6.0, 6.0, 6.0, 3.0, 3.0, 3.0, 3.0, 3.0, 4.0, 4.0, 4.0, 4.0, 4.0, 8.0, 8.0, 8.0, 8.0, 8.0, 7.0, 7.0, 7.0, 7.0, 7.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 4.0, 4.0, 4.0, 4.0, 4.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 6.0, 5.0, 5.0, 5.0, 5.0, 5.0, 4.0, 4.0, 4.0, 4.0, 4.0, 1.0, 1.0, 1.0, 1.0, 1.0, 7.0, 7.0, 7.0, 7.0, 7.0, 4.0, 4.0, 4.0, 4.0, 4.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 4.0, 4.0, 4.0, 4.0, 4.0, 7.0, 7.0, 7.0, 7.0, 7.0, 6.0, 6.0, 6.0, 6.0, 6.0, 4.0, 4.0, 4.0, 4.0, 4.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 7.0, 7.0, 7.0, 7.0, 7.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 6.0, 6.0, 6.0, 6.0, 6.0, 5.0, 5.0, 5.0, 5.0, 5.0, 4.0, 4.0, 4.0, 4.0, 4.0, 3.0, 3.0, 3.0, 3.0, 3.0, 4.0, 4.0, 4.0, 4.0, 4.0, 8.0, 8.0, 8.0, 8.0, 8.0, 6.0, 6.0, 6.0, 6.0, 6.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 6.0, 6.0, 6.0, 6.0, 6.0, 1.0, 1.0, 1.0, 1.0, 1.0, 5.0, 5.0, 5.0, 5.0, 5.0, 6.0, 6.0, 6.0, 6.0, 6.0, 5.0, 5.0, 5.0, 5.0, 5.0, 7.0, 7.0, 7.0, 7.0, 7.0, 5.0, 5.0, 5.0, 5.0, 5.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0]

This change introduces a llamafile_mixmul() API that allows tinyBLAS to speed up "Mixture of Expert" models. On my Threadripper, Mixtral's 8x7b F16 weights now process prompts 2x faster. I'm also seeing a 60 percent improvement with Mixtral 8x22b Q4_0. The same applies to Q8_0, which is also supported by tinyBLAS. MoE models spend the majority of their time inside MUL_MAT_ID rather than MUL_MAT, which is why llamafile_sgemm was not able to help them before. llamafile_mixmul works by decomposing the mixmul operation into sgemm calls.

lemmi · 2024-05-12T09:41:58Z

Here's a benchmark of an AMD V3C48 (a Zen 3 part) with Mistral 7B Instruct v0.2. (I had to throw out some code that used the X86_HAVE(F16C) check to make it compile though)

PR:

model	size	params	backend	threads	test	t/s
mistral 7B BF16	13.49 GiB	7.24 B	CPU	6	pp512	23.25 ± 0.17
mistral 7B BF16	13.49 GiB	7.24 B	CPU	6	tg128	3.04 ± 0.01
mistral 7B F16	13.49 GiB	7.24 B	CPU	6	pp512	23.10 ± 0.10
mistral 7B F16	13.49 GiB	7.24 B	CPU	6	tg128	3.02 ± 0.01
mistral 7B all F32	26.98 GiB	7.24 B	CPU	6	pp512	20.00 ± 0.06
mistral 7B all F32	26.98 GiB	7.24 B	CPU	6	tg128	1.52 ± 0.01

Without these changes, prompt processing for BF16 clocks in at about 11 t/s (see #7182), rest stays the same. Good improvement overall :)

(I'm still a bit confused as to why F16 performs so much better than BF16 without tinyblas and whether there is still something left on the table, but at least this way there is no compromise in using BF16 now)

ggml.c

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

jart · 2024-05-12T15:29:46Z

@lemmi where you're going to see the biggest changes here are running mixtral (rather than mistral) because moe models use MUL_MAT_ID (that's where they spend most of their clock cycles) and until this change we had no BLAS support whatsoever for MUL_MAT_ID. As for BF16 vs. F16 this change introduces tinyBLAS support for BF16 (which is a very recently introduced data type) so finally having BLAS-like performance for BF16 is naturally going to help it catch up with F16, and then surpass it on znver4.

ggerganov · 2024-05-12T15:39:21Z

Take a look at the failing CI run: https://github.com/ggerganov/llama.cpp/actions/runs/9052429493/job/24870086560?pr=6840

D:\a\llama.cpp\llama.cpp\sgemm.cpp(827,59): error C3861: 'MM256_SET_M128I': identifier not found [D:\a\llama.cpp\llama.cpp\build\ggml.vcxproj]

ggerganov · 2024-05-17T07:05:07Z

@jart I think the following patch should fix the CI:

diff --git a/ggml-impl.h b/ggml-impl.h
index d85b152b..85d3f23f 100644
--- a/ggml-impl.h
+++ b/ggml-impl.h
@@ -17,6 +17,9 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
 /**
  * Converts brain16 to float32.
  *
diff --git a/ggml-quants.c b/ggml-quants.c
index 00334c5f..3677b2db 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -22,9 +22,6 @@
 
 #define UNUSED GGML_UNUSED
 
-// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
-#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
-
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
 // multiply int8_t, add results pairwise twice
 static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {

jart force-pushed the moe branch 2 times, most recently from def794c to 828f3fe Compare April 23, 2024 05:44

jart force-pushed the moe branch from 828f3fe to 00fa7cd Compare April 23, 2024 17:49

jart force-pushed the moe branch from 00fa7cd to b0d0cdc Compare April 23, 2024 19:41

jart force-pushed the moe branch 2 times, most recently from 26ab943 to 89991a1 Compare April 25, 2024 00:56

jart force-pushed the moe branch from 89991a1 to bc1585b Compare April 25, 2024 01:52

jart force-pushed the moe branch 2 times, most recently from f1a134a to c34c472 Compare April 26, 2024 15:47

This was referenced Apr 26, 2024

Question Mozilla-Ocho/llamafile#363

Closed

AVX Q4_0 and Q8_0 sgemm #6891

Merged

ikawrakow mentioned this pull request May 3, 2024

Faster AVX2 prompt processing for k-quants and IQ4_XS Mozilla-Ocho/llamafile#394

Merged

jart mentioned this pull request May 3, 2024

Introduce bfloat16 support #6412

Merged

mofosyne added review complexity : medium Generally require more time to grok but manageable by beginner to medium expertise level enhancement New feature or request labels May 9, 2024

jart force-pushed the moe branch from c34c472 to 5b45ad4 Compare May 11, 2024 23:19

jart mentioned this pull request May 11, 2024

BF16 prompt processing has half the performance compared to F16 and F32 von AMD Ryzen Embedded V3000 (Zen 3) #7182

Open

jart force-pushed the moe branch from 5b45ad4 to 957670a Compare May 12, 2024 02:00

jart force-pushed the moe branch from 957670a to e717fec Compare May 12, 2024 02:44

jart force-pushed the moe branch from e717fec to e1c02a7 Compare May 12, 2024 03:20

ggerganov approved these changes May 12, 2024

View reviewed changes

ggml.c Outdated Show resolved Hide resolved

Update ggml.c

bb3a527

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

ggerganov added the merging soon Will merge soon unless anyone objects label May 15, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

llamafile : improve moe prompt eval speed on cpu #6840

llamafile : improve moe prompt eval speed on cpu #6840

jart commented Apr 23, 2024

hiepxanh commented Apr 23, 2024

USBhost commented Apr 23, 2024 •

edited

jart commented Apr 23, 2024

USBhost commented Apr 23, 2024 •

edited

jart commented Apr 23, 2024

jart commented Apr 24, 2024

jart commented Apr 25, 2024

ikawrakow commented Apr 26, 2024

jart commented Apr 26, 2024

ikawrakow commented Apr 26, 2024

jart commented Apr 26, 2024

github-actions bot commented Apr 26, 2024 •

edited

lemmi commented May 12, 2024

jart commented May 12, 2024

ggerganov commented May 12, 2024

ggerganov commented May 17, 2024

llamafile : improve moe prompt eval speed on cpu #6840

Are you sure you want to change the base?

llamafile : improve moe prompt eval speed on cpu #6840

Conversation

jart commented Apr 23, 2024

hiepxanh commented Apr 23, 2024

USBhost commented Apr 23, 2024 • edited

jart commented Apr 23, 2024

USBhost commented Apr 23, 2024 • edited

jart commented Apr 23, 2024

jart commented Apr 24, 2024

jart commented Apr 25, 2024

ikawrakow commented Apr 26, 2024

jart commented Apr 26, 2024

ikawrakow commented Apr 26, 2024

jart commented Apr 26, 2024

github-actions bot commented Apr 26, 2024 • edited

lemmi commented May 12, 2024

jart commented May 12, 2024

ggerganov commented May 12, 2024

ggerganov commented May 17, 2024

USBhost commented Apr 23, 2024 •

edited

USBhost commented Apr 23, 2024 •

edited

github-actions bot commented Apr 26, 2024 •

edited