Make GGML_SYCL_F16=ON the default by malsbat · Pull Request #23996 · ggml-org/llama.cpp

malsbat · 2026-06-01T21:23:07Z

Overview

The current default of GGML_SYCL_F16 is OFF. There are significant performance gains to be had during prompt processing by setting it ON: improvement on models measured averaged 2.43x (ranged from 1.48x to 3.41x) for prompt processing. Token generation measured no improvement (average of 1.00x, ranged from 0.97x to 1.02x).

At the same time, llama-perplexity showed no meaningful degradation on the wiki.test.raw, winogrande, or hellaswag benchmarks.

Additional information

Measurements below were captured on B70.

Performance measurements

Metric	Value
Average prefill speedup (F16/F32)	2.43x
Average decode speedup (F16/F32)	1.00x

Model	Task	Tokens	F32 (tok/s)	F16 (tok/s)	Speedup
DeepSeek-R1-Qwen-32B-Q4	pp	512	255.34 ±0.00	702.80 ±0.00	2.75x
DeepSeek-R1-Qwen-32B-Q4	pp	1024	269.73 ±0.00	850.26 ±0.00	3.15x
DeepSeek-R1-Qwen-32B-Q4	pp	2048	261.36 ±0.00	816.25 ±0.00	3.12x
DeepSeek-R1-Qwen-32B-Q4	pp	4096	237.66 ±0.00	625.70 ±0.00	2.63x
DeepSeek-R1-Qwen-32B-Q4	pp	8192	201.03 ±0.00	428.58 ±0.00	2.13x
DeepSeek-R1-Qwen-32B-Q4	tg	128	20.99 ±0.00	20.97 ±0.00	1.00x
DeepSeek-R1-Qwen-32B-Q4	tg	256	20.91 ±0.00	20.93 ±0.00	1.00x
DeepSeek-R1-Qwen-32B-Q4	tg	512	20.76 ±0.00	20.81 ±0.00	1.00x
DeepSeek-R1-Qwen-32B-Q4	tg	1024	20.52 ±0.00	20.48 ±0.00	1.00x
Gemma-2-9B	pp	512	861.32 ±0.00	1993.83 ±0.00	2.31x
Gemma-2-9B	pp	1024	884.10 ±0.00	2257.81 ±0.00	2.55x
Gemma-2-9B	pp	2048	842.96 ±0.00	1980.15 ±0.00	2.35x
Gemma-2-9B	pp	4096	720.86 ±0.00	1423.74 ±0.00	1.98x
Gemma-2-9B	pp	8192	566.30 ±0.00	929.83 ±0.00	1.64x
Gemma-2-9B	tg	128	58.58 ±0.00	58.79 ±0.00	1.00x
Gemma-2-9B	tg	256	58.35 ±0.00	58.46 ±0.00	1.00x
Gemma-2-9B	tg	512	58.10 ±0.00	58.19 ±0.00	1.00x
Gemma-2-9B	tg	1024	55.93 ±0.00	56.17 ±0.00	1.00x
Llama-3.1-8B-Q8	pp	512	1045.67 ±0.00	2745.04 ±0.00	2.63x
Llama-3.1-8B-Q8	pp	1024	1103.75 ±0.00	3091.26 ±0.00	2.80x
Llama-3.1-8B-Q8	pp	2048	1034.44 ±0.00	2621.40 ±0.00	2.53x
Llama-3.1-8B-Q8	pp	4096	890.18 ±0.00	1903.68 ±0.00	2.14x
Llama-3.1-8B-Q8	pp	8192	695.90 ±0.00	1209.49 ±0.00	1.74x
Llama-3.1-8B-Q8	tg	128	56.27 ±0.00	56.22 ±0.00	1.00x
Llama-3.1-8B-Q8	tg	256	56.30 ±0.00	56.23 ±0.00	1.00x
Llama-3.1-8B-Q8	tg	512	56.18 ±0.00	56.10 ±0.00	1.00x
Llama-3.1-8B-Q8	tg	1024	55.59 ±0.00	55.50 ±0.00	1.00x
Llama-3.2-3B	pp	512	2485.88 ±0.00	4973.16 ±0.00	2.00x
Llama-3.2-3B	pp	1024	2455.36 ±0.00	5407.97 ±0.00	2.20x
Llama-3.2-3B	pp	2048	2196.37 ±0.00	4455.35 ±0.00	2.03x
Llama-3.2-3B	pp	4096	1793.81 ±0.00	3118.76 ±0.00	1.74x
Llama-3.2-3B	pp	8192	1305.02 ±0.00	1933.71 ±0.00	1.48x
Llama-3.2-3B	tg	128	144.29 ±0.00	145.12 ±0.00	1.01x
Llama-3.2-3B	tg	256	144.03 ±0.00	144.28 ±0.00	1.00x
Llama-3.2-3B	tg	512	142.76 ±0.00	142.95 ±0.00	1.00x
Llama-3.2-3B	tg	1024	137.64 ±0.00	138.03 ±0.00	1.00x
Mistral-Nemo-12B	pp	512	702.92 ±0.00	1786.92 ±0.00	2.54x
Mistral-Nemo-12B	pp	1024	740.79 ±0.00	2064.76 ±0.00	2.79x
Mistral-Nemo-12B	pp	2048	705.70 ±0.00	1890.57 ±0.00	2.68x
Mistral-Nemo-12B	pp	4096	621.40 ±0.00	1396.77 ±0.00	2.25x
Mistral-Nemo-12B	pp	8192	501.28 ±0.00	915.55 ±0.00	1.83x
Mistral-Nemo-12B	tg	128	54.60 ±0.00	54.31 ±0.00	0.99x
Mistral-Nemo-12B	tg	256	54.46 ±0.00	53.95 ±0.00	0.99x
Mistral-Nemo-12B	tg	512	54.34 ±0.00	53.74 ±0.00	0.99x
Mistral-Nemo-12B	tg	1024	53.69 ±0.00	53.09 ±0.00	0.99x
Mistral-Small-24B	pp	512	360.64 ±0.00	1078.51 ±0.00	2.99x
Mistral-Small-24B	pp	1024	381.18 ±0.00	1308.88 ±0.00	3.43x
Mistral-Small-24B	pp	2048	377.68 ±0.00	1303.19 ±0.00	3.45x
Mistral-Small-24B	pp	4096	351.74 ±0.00	1045.46 ±0.00	2.97x
Mistral-Small-24B	pp	8192	309.36 ±0.00	753.30 ±0.00	2.44x
Mistral-Small-24B	tg	128	29.70 ±0.00	30.02 ±0.00	1.01x
Mistral-Small-24B	tg	256	29.63 ±0.00	29.92 ±0.00	1.01x
Mistral-Small-24B	tg	512	29.50 ±0.00	29.85 ±0.00	1.01x
Mistral-Small-24B	tg	1024	29.30 ±0.00	29.63 ±0.00	1.01x
Phi-3.5-mini-3.8B	pp	512	1853.10 ±0.00	4075.13 ±0.00	2.20x
Phi-3.5-mini-3.8B	pp	1024	1843.66 ±0.00	4235.75 ±0.00	2.30x
Phi-3.5-mini-3.8B	pp	2048	1693.64 ±0.00	3417.81 ±0.00	2.02x
Phi-3.5-mini-3.8B	pp	4096	1367.47 ±0.00	2397.30 ±0.00	1.75x
Phi-3.5-mini-3.8B	pp	8192	992.83 ±0.00	1471.76 ±0.00	1.48x
Phi-3.5-mini-3.8B	tg	128	134.95 ±0.00	135.52 ±0.00	1.00x
Phi-3.5-mini-3.8B	tg	256	134.57 ±0.00	134.70 ±0.00	1.00x
Phi-3.5-mini-3.8B	tg	512	130.86 ±0.00	131.12 ±0.00	1.00x
Phi-3.5-mini-3.8B	tg	1024	124.89 ±0.00	125.02 ±0.00	1.00x
Qwen2.5-14B-Q4	pp	512	565.78 ±0.00	1393.04 ±0.00	2.46x
Qwen2.5-14B-Q4	pp	1024	595.27 ±0.00	1595.79 ±0.00	2.68x
Qwen2.5-14B-Q4	pp	2048	560.66 ±0.00	1433.56 ±0.00	2.56x
Qwen2.5-14B-Q4	pp	4096	482.25 ±0.00	1023.94 ±0.00	2.12x
Qwen2.5-14B-Q4	pp	8192	378.14 ±0.00	655.40 ±0.00	1.73x
Qwen2.5-14B-Q4	tg	128	43.64 ±0.00	43.55 ±0.00	1.00x
Qwen2.5-14B-Q4	tg	256	43.43 ±0.00	43.37 ±0.00	1.00x
Qwen2.5-14B-Q4	tg	512	43.16 ±0.00	43.07 ±0.00	1.00x
Qwen2.5-14B-Q4	tg	1024	42.25 ±0.00	42.19 ±0.00	1.00x
Qwen2.5-14B-Q8	pp	512	557.42 ±0.00	1461.46 ±0.00	2.62x
Qwen2.5-14B-Q8	pp	1024	590.94 ±0.00	1645.93 ±0.00	2.79x
Qwen2.5-14B-Q8	pp	2048	559.19 ±0.00	1432.49 ±0.00	2.56x
Qwen2.5-14B-Q8	pp	4096	481.75 ±0.00	1018.36 ±0.00	2.11x
Qwen2.5-14B-Q8	pp	8192	377.23 ±0.00	647.09 ±0.00	1.72x
Qwen2.5-14B-Q8	tg	128	28.59 ±0.00	28.61 ±0.00	1.00x
Qwen2.5-14B-Q8	tg	256	28.57 ±0.00	28.61 ±0.00	1.00x
Qwen2.5-14B-Q8	tg	512	28.43 ±0.00	28.46 ±0.00	1.00x
Qwen2.5-14B-Q8	tg	1024	28.04 ±0.00	28.06 ±0.00	1.00x
Qwen2.5-32B-Q4	pp	512	250.16 ±0.00	703.77 ±0.00	2.81x
Qwen2.5-32B-Q4	pp	1024	262.57 ±0.00	848.11 ±0.00	3.23x
Qwen2.5-32B-Q4	pp	2048	255.08 ±0.00	810.17 ±0.00	3.18x
Qwen2.5-32B-Q4	pp	4096	232.17 ±0.00	624.48 ±0.00	2.69x
Qwen2.5-32B-Q4	pp	8192	196.77 ±0.00	428.02 ±0.00	2.18x
Qwen2.5-32B-Q4	tg	128	20.58 ±0.00	20.94 ±0.00	1.02x
Qwen2.5-32B-Q4	tg	256	20.53 ±0.00	20.86 ±0.00	1.02x
Qwen2.5-32B-Q4	tg	512	20.43 ±0.00	20.73 ±0.00	1.01x
Qwen2.5-32B-Q4	tg	1024	20.15 ±0.00	20.46 ±0.00	1.02x
Qwen2.5-32B-Q6	pp	512	252.12 ±0.00	756.21 ±0.00	3.00x
Qwen2.5-32B-Q6	pp	1024	267.10 ±0.00	858.42 ±0.00	3.21x
Qwen2.5-32B-Q6	pp	2048	258.91 ±0.00	809.55 ±0.00	3.13x
Qwen2.5-32B-Q6	pp	4096	236.23 ±0.00	619.63 ±0.00	2.62x
Qwen2.5-32B-Q6	pp	8192	200.27 ±0.00	423.10 ±0.00	2.11x
Qwen2.5-32B-Q6	tg	128	12.83 ±0.00	12.49 ±0.00	0.97x
Qwen2.5-32B-Q6	tg	256	12.76 ±0.00	12.44 ±0.00	0.97x
Qwen2.5-32B-Q6	tg	512	12.76 ±0.00	12.41 ±0.00	0.97x
Qwen2.5-32B-Q6	tg	1024	12.66 ±0.00	12.32 ±0.00	0.97x
Qwen2.5-7B	pp	512	1142.70 ±0.00	2918.44 ±0.00	2.55x
Qwen2.5-7B	pp	1024	1213.82 ±0.00	3246.44 ±0.00	2.67x
Qwen2.5-7B	pp	2048	1165.09 ±0.00	3090.48 ±0.00	2.65x
Qwen2.5-7B	pp	4096	1021.99 ±0.00	2300.18 ±0.00	2.25x
Qwen2.5-7B	pp	8192	821.50 ±0.00	1505.63 ±0.00	1.83x
Qwen2.5-7B	tg	128	82.13 ±0.00	82.39 ±0.00	1.00x
Qwen2.5-7B	tg	256	81.62 ±0.00	82.00 ±0.00	1.00x
Qwen2.5-7B	tg	512	80.06 ±0.00	80.47 ±0.00	1.01x
Qwen2.5-7B	tg	1024	78.68 ±0.00	79.11 ±0.00	1.01x
Qwen3-8B	pp	512	1040.88 ±0.00	2483.45 ±0.00	2.39x
Qwen3-8B	pp	1024	1060.96 ±0.00	2812.53 ±0.00	2.65x
Qwen3-8B	pp	2048	1010.09 ±0.00	2409.60 ±0.00	2.39x
Qwen3-8B	pp	4096	857.38 ±0.00	1722.44 ±0.00	2.01x
Qwen3-8B	pp	8192	658.54 ±0.00	1095.43 ±0.00	1.66x
Qwen3-8B	tg	128	76.29 ±0.00	76.65 ±0.00	1.00x
Qwen3-8B	tg	256	76.07 ±0.00	76.53 ±0.00	1.01x
Qwen3-8B	tg	512	75.89 ±0.00	76.19 ±0.00	1.00x
Qwen3-8B	tg	1024	74.66 ±0.00	75.13 ±0.00	1.01x
Qwen3.5-9B-Q4	pp	512	1017.20 ±0.00	2390.18 ±0.00	2.35x
Qwen3.5-9B-Q4	pp	1024	1047.11 ±0.00	2727.16 ±0.00	2.60x
Qwen3.5-9B-Q4	pp	2048	1066.86 ±0.00	2749.60 ±0.00	2.58x
Qwen3.5-9B-Q4	pp	4096	1020.11 ±0.00	2504.86 ±0.00	2.46x
Qwen3.5-9B-Q4	pp	8192	943.54 ±0.00	2112.99 ±0.00	2.24x
Qwen3.5-9B-Q4	tg	128	65.74 ±0.00	65.56 ±0.00	1.00x
Qwen3.5-9B-Q4	tg	256	65.47 ±0.00	65.27 ±0.00	1.00x
Qwen3.5-9B-Q4	tg	512	65.42 ±0.00	65.15 ±0.00	1.00x
Qwen3.5-9B-Q4	tg	1024	65.30 ±0.00	65.08 ±0.00	1.00x

Perplexity measurements

hellaswag

Model	Metric	F32 Score	F16 Score	Delta (F16-F32)	Ratio (F16/F32)
DeepSeek-R1-Qwen-32B-Q4	hellaswag	80.0000	80.0000	+0.0000	1.0000
Gemma-2-9B	hellaswag	78.5000	78.5000	+0.0000	1.0000
Llama-3.1-8B-Q8	hellaswag	78.5000	78.5000	+0.0000	1.0000
Llama-3.2-3B	hellaswag	73.0000	73.0000	+0.0000	1.0000
Mistral-Nemo-12B	hellaswag	81.5000	81.5000	+0.0000	1.0000
Mistral-Small-24B	hellaswag	82.0000	82.0000	+0.0000	1.0000
Phi-3.5-mini-3.8B	hellaswag	76.5000	76.7500	+0.2500	1.0033
Qwen2.5-14B-Q4	hellaswag	82.7500	82.7500	+0.0000	1.0000
Qwen2.5-14B-Q8	hellaswag	82.7500	82.7500	+0.0000	1.0000
Qwen2.5-32B-Q4	hellaswag	84.7500	84.7500	+0.0000	1.0000
Qwen2.5-32B-Q6	hellaswag	84.7500	84.7500	+0.0000	1.0000
Qwen2.5-7B	hellaswag	80.0000	80.0000	+0.0000	1.0000
Qwen3-8B	hellaswag	73.7500	73.7500	+0.0000	1.0000
Qwen3.5-9B-Q4	hellaswag	78.0000	78.0000	+0.0000	1.0000

wiki.test.raw

Model	Metric	F32 PPL	F16 PPL	Delta (F16-F32)	Ratio (F16/F32)
DeepSeek-R1-Qwen-32B-Q4	perplexity	7.1700	7.1696	-0.0004	0.9999
Gemma-2-9B	perplexity	8.7995	8.7993	-0.0002	1.0000
Llama-3.1-8B-Q8	perplexity	7.3245	7.3245	+0.0000	1.0000
Llama-3.2-3B	perplexity	10.7252	10.7252	+0.0000	1.0000
Mistral-Nemo-12B	perplexity	6.4357	6.4357	+0.0000	1.0000
Mistral-Small-24B	perplexity	5.9889	5.9888	-0.0001	1.0000
Phi-3.5-mini-3.8B	perplexity	6.6916	6.6916	+0.0000	1.0000
Qwen2.5-14B-Q4	perplexity	6.1145	6.1145	+0.0000	1.0000
Qwen2.5-14B-Q8	perplexity	5.9765	5.9766	+0.0001	1.0000
Qwen2.5-32B-Q4	perplexity	5.6140	5.6142	+0.0002	1.0000
Qwen2.5-32B-Q6	perplexity	5.5257	5.5257	+0.0000	1.0000
Qwen2.5-7B	perplexity	7.9973	7.9974	+0.0001	1.0000
Qwen3-8B	perplexity	10.4663	10.4663	+0.0000	1.0000
Qwen3.5-9B-Q4	perplexity	8.3137	8.3137	+0.0000	1.0000

winogrande

Model	Metric	F32 Score	F16 Score	Delta (F16-F32)	Ratio (F16/F32)
DeepSeek-R1-Qwen-32B-Q4	winogrande	75.5328	75.5328	+0.0000	1.0000
Gemma-2-9B	winogrande	76.4799	76.4009	-0.0790	0.9990
Llama-3.1-8B-Q8	winogrande	73.5596	73.5596	+0.0000	1.0000
Llama-3.2-3B	winogrande	68.1926	68.1926	+0.0000	1.0000
Mistral-Nemo-12B	winogrande	76.9534	76.9534	+0.0000	1.0000
Mistral-Small-24B	winogrande	79.2423	79.2423	+0.0000	1.0000
Phi-3.5-mini-3.8B	winogrande	73.7964	73.7964	+0.0000	1.0000
Qwen2.5-14B-Q4	winogrande	74.1910	74.0331	-0.1579	0.9979
Qwen2.5-14B-Q8	winogrande	73.8753	73.8753	+0.0000	1.0000
Qwen2.5-32B-Q4	winogrande	72.5335	72.4546	-0.0789	0.9989
Qwen2.5-32B-Q6	winogrande	72.9282	72.8493	-0.0789	0.9989
Qwen2.5-7B	winogrande	71.0339	71.0339	+0.0000	1.0000
Qwen3-8B	winogrande	68.5083	68.4294	-0.0789	0.9988
Qwen3.5-9B-Q4	winogrande	71.7443	71.7443	+0.0000	1.0000

Requirements

I have read and agree with the contributing guidelines
AI usage disclosure: NO

Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

ggml-gh-bot · 2026-06-01T21:28:15Z

Hi @malsbat, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

arthw

SYCL support fp32 and fp16 for two reasons:

fp32 will provide better accuracy than fp16.
fp32 show better performance than fp16 in some cases.

I agree to recommend to use fp16 as default.

But not change the default value of parameter GGML_SYCL_F16 and the default values in script: That will impact existed CI in all user development.

malsbat · 2026-06-02T16:14:24Z

@arthw, thanks for reviewing. My measurements are quite different than what you're seeing - I see better performance with fp16 and no meaningful degradation in quality. Can you tell me which models and devices you're seeing better results with fp32? I'd like to add these to my local benchmarks if possible.

arthw · 2026-06-03T12:25:38Z

@malsbat
We collect the performance data and saved in #23313.

We can see the PP of FP16 is about double performance of FP32, but the TG is similar (FP32 is little bigger than FP16 in more cases).

I think FP16 has half bandwidth than FP32 in PP, for TG FP32 code path use fp16 data type in computing in fact.

I'm considering to merge FP16 and FP32 code path to one in the future. Like PP is using FP16 and TG is using FP32.

Anyway, I think recommend FP16 is good idea.
We only need to change the recommended mode in SYCL.md.
To avoid to impact the existed CI/docker/delivery, I suggest not changing the default value and script for it.
The dockerfile, it's OK to set FP16=on.

Thank you!

F16 remains explictly set for example and Dockerfile builds. Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

malsbat · 2026-06-03T16:35:36Z

Thanks @arthw, I may have misunderstood your earlier comment. I've updated the PR to leave F32 the default in CMakeLists.txt and explicitly set F16=ON in the other locations.

I did notice that https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh#L113 explicitly sets F16=ON, so I'm unsure of your comment about impacting existing CI. Please let me know if this PR matches what you intended.

The main change I want is to the Dockerfile so that end-users get the best performance when using the prebuilt images.

arthw · 2026-06-08T14:20:23Z

@malsbat
I see your update.

ci/run.sh - In fact, SYCL backend only CI is only building, not execute this script.
I will execute the CI locally, but not using this script.

build.sh and win-build-sycl.bat: I hope keep the default behavior (fp32), so that the existed CI (different users locally) won't be impacted.

I'm considering to merge fp32 and fp16 code paths in different function/OPs, to get the best performance.
The flag fp32 and fp16 will disappear in code in the future.

So we won't need to modify more on this issue now.

Thank you!

malsbat · 2026-06-08T17:10:42Z

Should I close this PR then?

arthw · 2026-06-09T07:34:04Z

No! :)
Only rollback: build.sh and win-build-sycl.bat.

Thank you!

malsbat · 2026-06-09T22:51:09Z

Thanks for clarifying, I will update this PR shortly.

Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

malsbat · 2026-06-09T23:00:21Z

Updated, please review when you can

arthw

It's good job!

Thank you!

ggerganov · 2026-06-15T07:00:34Z

@arthw Are you sure you want to make this change - F16 accumulators will definitely overflow in some cases, so generally not recommended.

arthw · 2026-06-15T08:01:38Z

@arthw Are you sure you want to make this change - F16 accumulators will definitely overflow in some cases, so generally not recommended.

@ggerganov
The overflow in FP16 is present on some cases.
But there are good results of FP16 in many cases. Please refer to #23313.

I think FP16 will be chosen by most user if there is no such overflow issue.
We didn't see the abnormal output in FP16 case for a long time.

This PR is used to change to FP16 in docker building and recommend FP16 in guide.
User can switch to FP32 in any time.

In same time, the docker user will feedback the overflow issue as docker use FP16 as default.
It will help us to collect the data of FP16.
The impact will be limited.

Another issue, I'm considering to merge FP32 and FP16 in SYCL backend: choose the performance better part in kernel.
But as your mentioned, there is risk to remove FP32 in some OPs.
I will think about it.

Thank you!

* Add -cl-fp32-correctly-rounded-divide-sqrt to F16=ON builds Signed-off-by: Todd Malsbary <todd.malsbary@intel.com> * Make GGML_SYCL_F16=ON the default Signed-off-by: Todd Malsbary <todd.malsbary@intel.com> * Leave F32 the default F16 remains explictly set for example and Dockerfile builds. Signed-off-by: Todd Malsbary <todd.malsbary@intel.com> * Revert changes to examples/sycl/build scripts Signed-off-by: Todd Malsbary <todd.malsbary@intel.com> --------- Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

malsbat added 2 commits May 13, 2026 23:32

Add -cl-fp32-correctly-rounded-divide-sqrt to F16=ON builds

c1ec115

Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

Make GGML_SYCL_F16=ON the default

5b77745

Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

malsbat requested a review from ngxson as a code owner June 1, 2026 21:23

Merge remote-tracking branch 'ggml-org/master' into sycl-f16-default

060bab3

Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

arthw reviewed Jun 2, 2026

View reviewed changes

Leave F32 the default

60c2372

F16 remains explictly set for example and Dockerfile builds. Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

Revert changes to examples/sycl/build scripts

a495ce0

Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>

arthw approved these changes Jun 10, 2026

View reviewed changes

arthw added the merge ready A maintainer can use this label to indicate that they consider the changes final and ready to merge. label Jun 11, 2026

ggerganov merged commit 4196b47 into ggml-org:master Jun 16, 2026
3 checks passed

Uh oh!

Conversation

malsbat commented Jun 1, 2026

Overview

Additional information

Performance measurements

Perplexity measurements

hellaswag

wiki.test.raw

winogrande

Requirements

Uh oh!

ggml-gh-bot Bot commented Jun 1, 2026

Uh oh!

arthw left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

malsbat commented Jun 2, 2026

Uh oh!

arthw commented Jun 3, 2026

Uh oh!

malsbat commented Jun 3, 2026

Uh oh!

arthw commented Jun 8, 2026

Uh oh!

malsbat commented Jun 8, 2026

Uh oh!

arthw commented Jun 9, 2026

Uh oh!

malsbat commented Jun 9, 2026

Uh oh!

malsbat commented Jun 9, 2026

Uh oh!

arthw left a comment

Choose a reason for hiding this comment

Uh oh!

ggerganov commented Jun 15, 2026

Uh oh!

arthw commented Jun 15, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

arthw left a comment •

edited

Loading