ggml : remove bit shuffling

[ggerganov]

Implementation of #1241

Avoid unnecessary bit shuffling by packing the quants in a better way.
Requires model re-quantization

• Q4_0

  • quantize
  • dequantize
  • dot ARM NEON
  • dot AVX

• Q4_1

  • quantize
  • dequantize
  • dot ARM NEON
  • dot AVX

• Q5_0

  • quantize
  • dequantize
  • dot ARM NEON
  • dot AVX
  • dot WASM SIMD

• Q5_1

  • quantize
  • dequantize
  • dot ARM NEON
  • dot AVX
  • dot WASM SIMD

---

New timings:

Model	Measure	F16	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0
7B	ms/tok @ 4th	127	49	56	89	92	74
7B	ms/tok @ 8th	120	44	52	49	52	70
13B	ms/tok @ 4th	261*	91	103	173	177	139
13B	ms/tok @ 8th	316*	81	95	103	113	134

• these numbers vary a lot since the model is on the 32GB limit of my MacBook

Old timings:

Model	Measure	F16	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0
7B	ms/tok @ 4th	128	56	61	91	95	75
7B	ms/tok @ 8th	128	47	55	53	59	75
13B	ms/tok @ 4th	239	104	113	176	185	141
13B	ms/tok @ 8th	240	85	99	108	117	147

overall, all these numbers seem to have about +/- 10% variablility from run to run. not ideal benchmark, but not sure what else to do

[ggerganov]

Unfortunately, Q4_2 does not fit into this pattern, unless we introduce a Q8_2 block with size of 16 instead of 32 elements.
Any suggestions how to proceed? Maybe drop Q4_2 support?

[sw]

A few remarks:

• quantize_row_q4_1_reference is broken, it's missing the i*qk in x[...]
• Q5 interleaves the nibbles but not the single MSBs, that could be confusing. The scalar implementation seems to be broken, maybe it's because of that?
• I still think this could be done in a non-breaking manner for Q4, by simply changing the order of Q8

[ggerganov]

@sw

I still think this could be done in a non-breaking manner for Q4, by simply changing the order of Q8

Yes, I'm still hesitating. But I think Q8 quantization will be sub-optimal this way.
Not much, but still.

[ggerganov]

Q5 interleaves the nibbles but not the single MSBs, that could be confusing. The scalar implementation seems to be broken, maybe it's because of that?

Somehow perplexity computation with Q5_0 + cuBLAS is currently broken and I don't see the error.
It works on M1.

Edit: fixed

[hmih]

Is there enough bloat caused by supporting the old models that a header change to allow loading of both the new and old versions would be an imposition? That would keep old model compatibility and allow new quants to benefit from the speedup without any impact on end users.

"End users" here are not state bureaucrats with IE8, but adventurous devs who are involved with an experimental approach to a new technology. Breakage is the name of the game. It takes a minute to cleanup and rerun the scripts. For my models I prefer minimal and fast. If anything I would like to have the possibility to break compatibility for the sake of performance and size.

[LostRuins]

@hmih With all due respect, I know this project is all about optimization and experimentation. I am just suggesting that since this is such a major change, helping others identify older and newer models elsewhere would be very useful, since there is already a thriving ggml ecosystem beyond this repo.

I understand that reconverting the models for those familiar with this project is easy enough - but I just think a 1 line change for the file magic, or simply incrementing the existing version header by 1, something that requires minimal effort, would make future maintenance and identifying these new formats easier.

The in-place swap suggested by @digiwombat would be even better if possible.

[digiwombat]

"End users" here are not state bureaucrats with IE8

Firstly, I agree very much in spirit with your statement and the speedup the new version will bring. I also would offer that it is fairly early in the lifecycle for this project and that is an argument in favor of just putting through the breaking change and letting it be.

On the other side, ggml (especially via llama.cpp) is being used pretty widely and I would ask that a thought be spared for the maintainers of supporting projects like @LostRuins with Koboldcpp or the fine folks doing llama-cpp-python (and downstream of it, oobabooga's webui), among others who will likely bear the brunt of user confusion on these issues. It will cause a burden in a wider radius that the user-facing software people don't have a lot of control over since they are generally not in control of the model repos to make the updates themselves.

That's all. Just wanted to toss out a bit of an explanation since the nature of the users was raised. I think the repos for front-end projects may see the target userbase for their projects much differently than core llama.cpp, generally speaking.

[Green-Sky]

llama.cpp/ggml.h

Line 191 in 0e48eb6

#define GGML_FILE_VERSION 1

can we increment this value by 1 ?

edit: oh, it was all in the llama.h/.cpp

[sw]

can we increment this value by 1 ?

That would make the unaffected formats incompatible - F16, Q8. The clean way would be to define new formats Q4_4, Q4_5, etc. But that gets unwieldy quickly.

[LostRuins]

@sw doesn't have to be though, during loading exceptions can be added in llama.cpp to treat old f16 and q8 format with either file versions 1 or 2 as forwards compatible.

[ggerganov]

Close in favor of #1405

[ProfessorSparrs]

great, I finally compiled it on my pc(no avx2 support) AND with cuda support. And this change makes none of my models load :(. I dont know how to quantize things, Ive read a lot about it and I doubt I even have the PC-resources to do it.

[Green-Sky]

@ProfessorSparrs if you have the f16 files, qnantizing is very easy and WAY less recource intensive than running the model. :) (check the quantize executable)