ggml-cuda : add rope f16, restore performance with parallel decoding

[slaren]

With this change, all the offloaded leafs are assumed to be inputs and copied to VRAM. This allows offloading KQ_mask and simplifies the logic for offloading KQ_pos. Replacing a conditional in the RoPE kernel with a template parameter as suggested by @JohannesGaessler also improves performance.

Performance compared to master:

Device 0: NVIDIA GeForce RTX 3090 Ti, compute capability 8.6

model	size	backend	test	master t/s	PR t/s	speedup
llama 7B mostly Q2_K	2.63 GiB	CUDA	pp 512	1805.63 ± 10.64	1828.37 ± 5.60	1.01259
llama 7B mostly Q3_K - Small	2.75 GiB	CUDA	pp 512	1907.34 ± 5.34	1932.17 ± 5.03	1.01301
llama 7B mostly Q3_K - Medium	3.07 GiB	CUDA	pp 512	1997.51 ± 2.89	2017.52 ± 3.51	1.01001
llama 7B mostly Q3_K - Large	3.35 GiB	CUDA	pp 512	1928.15 ± 4.06	1951.13 ± 4.13	1.01191
llama 7B mostly Q4_0	3.56 GiB	CUDA	pp 512	2404.71 ± 3.79	2439.23 ± 5.24	1.01435
llama 7B mostly Q4_K - Small	3.59 GiB	CUDA	pp 512	2202.34 ± 3.30	2226.35 ± 5.70	1.01090
llama 7B mostly Q4_K - Medium	3.80 GiB	CUDA	pp 512	2207.27 ± 3.36	2229.99 ± 11.03	1.01029
llama 7B mostly Q4_1	3.95 GiB	CUDA	pp 512	2102.01 ± 3.47	2123.36 ± 3.03	1.01015
llama 7B mostly Q5_0	4.33 GiB	CUDA	pp 512	2172.19 ± 4.07	2202.96 ± 5.92	1.01416
llama 7B mostly Q5_K - Small	4.33 GiB	CUDA	pp 512	2035.79 ± 4.42	2059.83 ± 5.22	1.01180
llama 7B mostly Q5_K - Medium	4.45 GiB	CUDA	pp 512	2057.20 ± 4.03	2065.73 ± 24.31	1.00414
llama 7B mostly Q5_1	4.72 GiB	CUDA	pp 512	1931.67 ± 2.74	1948.27 ± 4.01	1.00859
llama 7B mostly Q6_K	5.15 GiB	CUDA	pp 512	2107.53 ± 1.66	2133.80 ± 7.40	1.01246
llama 7B mostly Q8_0	6.67 GiB	CUDA	pp 512	2353.88 ± 3.43	2386.22 ± 9.66	1.01373
llama 7B mostly F16	12.55 GiB	CUDA	pp 512	1659.42 ± 3.10	1679.42 ± 2.03	1.01205
llama 7B mostly Q2_K	2.63 GiB	CUDA	tg 128	105.91 ± 0.33	105.08 ± 1.66	0.99216
llama 7B mostly Q3_K - Small	2.75 GiB	CUDA	tg 128	101.85 ± 0.14	100.75 ± 1.97	0.98919
llama 7B mostly Q3_K - Medium	3.07 GiB	CUDA	tg 128	108.07 ± 0.30	107.95 ± 0.90	0.99888
llama 7B mostly Q3_K - Large	3.35 GiB	CUDA	tg 128	105.34 ± 0.25	105.65 ± 0.36	1.00294
llama 7B mostly Q4_0	3.56 GiB	CUDA	tg 128	131.00 ± 0.30	130.90 ± 0.85	0.99923
llama 7B mostly Q4_K - Small	3.59 GiB	CUDA	tg 128	119.67 ± 0.23	120.00 ± 0.42	1.00275
llama 7B mostly Q4_K - Medium	3.80 GiB	CUDA	tg 128	114.40 ± 0.14	114.15 ± 0.54	0.99781
llama 7B mostly Q4_1	3.95 GiB	CUDA	tg 128	124.94 ± 0.04	124.71 ± 0.67	0.99815
llama 7B mostly Q5_0	4.33 GiB	CUDA	tg 128	112.05 ± 0.43	111.67 ± 0.34	0.99660
llama 7B mostly Q5_K - Small	4.33 GiB	CUDA	tg 128	110.85 ± 0.11	111.08 ± 0.69	1.00207
llama 7B mostly Q5_K - Medium	4.45 GiB	CUDA	tg 128	107.38 ± 0.06	106.77 ± 0.57	0.99431
llama 7B mostly Q5_1	4.72 GiB	CUDA	tg 128	108.43 ± 0.23	108.22 ± 0.78	0.99806
llama 7B mostly Q6_K	5.15 GiB	CUDA	tg 128	92.17 ± 0.07	91.27 ± 0.35	0.99023
llama 7B mostly Q8_0	6.67 GiB	CUDA	tg 128	88.62 ± 0.04	87.79 ± 0.48	0.99063
llama 7B mostly F16	12.55 GiB	CUDA	tg 128	56.14 ± 0.07	55.90 ± 0.16	0.99572

[slaren]

This is to avoid creating a new leaf with ggml_cpy(.., ggml_new_tensor(..)).

[ggerganov]

After this change, the lora test is failing:

make -j && ./bin/perplexity --model ../tmp/mnt/models/open-llama/3B-v2/ggml-model-q8_0.gguf -f ../tmp/mnt/models/shakespeare/shakespeare.txt --lora ../tmp/mnt/models/open-llama/3B-v2/lora/ggml-adapter-model.bin  --lora-base ../tmp/mnt/models/open-llama/3B-v2/ggml-model-f16.gguf -c 128 -b 128 --chunks 2
Segmentation fault: 11

It crashes in ggml_build_forward_expand() for the Vcur cpy:

llama.cpp/llama.cpp

Line 2831 in 2f3a46f

ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));

[slaren]

Yeah, this change to ggml_cpy is not going to work. Maybe we could add an ggml_cont_reshape or just ggml_cont_4d that does the same thing?

Such as:

struct ggml_tensor * ggml_cont_4d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
        int64_t               ne0,
        int64_t               ne1,
        int64_t               ne2,
        int64_t               ne3) {
    GGML_ASSERT(ggml_nelements(a) == (ne0*ne1*ne2*ne3));

    bool is_node = false;

    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
    ggml_format_name(result, "%s (cont)", a->name);

    result->op   = GGML_OP_CONT;
    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
    result->src[0] = a;

    return result;
}

Then replace:

cur = ggml_cpy(ctx0,
        KQV_merged,
        ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens));

with

cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);

[ggerganov]

I'm testing this branch with the following command to fill-up the context and the results become incoherent after the first rope shift:

make -j && ./bin/main -m ../models/llama-7b/ggml-model-q8_0.gguf -p "I believe the meaning of life is" --ignore-eos -c 256 -n -1 -t 8 -ngl 35

It works OK with -ngl 34.
The following patch helps to see when the K-cache shifts occur, for easier debugging:

diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 1ed543c..2bf7628 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -502,6 +502,8 @@ int main(int argc, char ** argv) {
                 const int n_left    = n_past - params.n_keep - 1;
                 const int n_discard = n_left/2;
 
+                printf("\n\n XXXXXXXXXXXXXXXXXXXXXX\n\n");
+
                 LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);

llm_load_tensors: ggml ctx size =    0,09 MB
llm_load_tensors: using CUDA for GPU acceleration
llm_load_tensors: mem required  =  132,91 MB (+  128,00 MB per state)
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloading v cache to GPU
llm_load_tensors: offloading k cache to GPU
llm_load_tensors: offloaded 35/35 layers to GPU
llm_load_tensors: VRAM used: 6824 MB
...................................................................................................
llama_new_context_with_model: kv self size  =  128,00 MB
llama_new_context_with_model: compute buffer total size =   36,72 MB
llama_new_context_with_model: VRAM scratch buffer: 35,25 MB

system_info: n_threads = 8 / 64 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | 
sampling: repeat_last_n = 64, repeat_penalty = 1,100000, presence_penalty = 0,000000, frequency_penalty = 0,000000, top_k = 40, tfs_z = 1,000000, top_p = 0,950000, typical_p = 1,000000, temp = 0,800000, mirostat = 0, mirostat_lr = 0,100000, mirostat_ent = 5,000000
generate: n_ctx = 256, n_batch = 512, n_predict = -1, n_keep = 0


 I believe the meaning of life is to love and serve others. The ultimate goal is becoming a servant-leader, and the best way to do that is through service and action.
As a young child in Vietnam, I had many friends who were not as fortunate as I was because of our war situation. In my teenage years, this realization made me feel guilty for being one of the lucky ones. My family fled to the United States when I was 14 years old and I’ve been living in the USA ever since. As a refugee myself, I understand the struggles of other refugees better than others because of my personal experience.
Teaching English is my passion, and it’s what keeps me inspired to serve. I started volunteering for a local nonprofit organization teaching English as a second language in 2015, and it has been the most rewarding and fulfilling part of my life since then. It helps me feel happy, proud, and productive!
My experience with refugees is not limited to students from Vietnam only; I have taught students from different countries including: Afghanistan, Burma (Myanmar), Bhutan, Laos,

 XXXXXXXXXXXXXXXXXXXXXX

 I love teaching English as a volunteer for nearly three years in the Portland Learning Center. As an ESL teacher at the same center since 2015; it has enriched my life and career-wise, even more than ever! Every time when I see them succeeding and improving their lives is very important to me teaching English as a volunteering for over five years ago – some places such as Sudanese, but also Syria, Iraq, Somalia, Congo, Afghanistan, Iran, Pakistan, Nepal, Vietnam, Eritia (Laos, Mynamia, Russia

 XXXXXXXXXXXXXXXXXXXXXX

, have learn about culture and have helped students. I’s now. The center in Seattle with a volunteer. My job is very same time as 16. When i, given me with so much I am able to my students are improving their lives. Thankfully!
and progressing, and also; because they improve my lives. Now.
I, I have as well. Thailand, Somalia, Kenya, Burma, Sudan, Sri Laba, and Sudh, Syria, Bhutan, Nepali, and Iraq and China, India, Tibet, Y

 XXXXXXXXXXXXXXXXXXXXXX

v I, their 1 have also been working as a year, and have a year-I’ve. I as.
Bes. I have also am I the, am happy for me in. Thaning.
Kathing and with a my English is much a. Also, and, I am volunteered at, soy, a have worked at afr.
Sudan, and Rwia, Eritisnaa and I have worked for. The work for 10, Bhutan.
Viety. e. and Vietnand I have and

 XXXXXXXXXXXXXXXXXXXXXX

 5, I the the last 6 to as an the same that they have work in, a to be a and as a
BH and the 3 and my, an as the
My havee work have tto be.
I am I, I
The are work. asI and work,
I have a I am with the ean and India, am
I am and in India. have I’ve worked an me and
have I
I am my have a, I have 1 I, am. am, I have.
I be the
am have I,

 XXXXXXXXXXXXXXXXXXXXXX

 in
inH, and work, asI work
The as. I have
be the last year,
I I, I h1, last
5,
I. I the work. The same. an, be the. I am a have the
My 6,be the,
to 7.

[slaren]

Should be fixed now. Unfortunately, the fix also makes the copies of K_pos, K_mask and K_shift synchronous, because they may be overwritten by other operations once the computation start (probably the get_rows, which still runs on the CPU). It shouldn't affect performance too much, but it should be fixed in the future.