SYCL bug: DeepSeek-V2-Lite-Chat-Q4_K_M does not work as expected

[aubreyli]

Name and Version

root@alc-ai:/home/aubrey/work/llama-gpu# ./build/bin/llama-cli --version
version: 4887 (8fcb563)
built with Intel(R) oneAPI DPC++/C++ Compiler 2025.0.4 (2025.0.4.20241205) for x86_64-unknown-linux-gnu

Operating systems

No response

Which llama.cpp modules do you know to be affected?

No response

Command line

 ./build/bin/llama-cli -m /srv/models/DeepSeek-V2-Lite-Chat-Q4_K_M/DeepSeek-V2-Lite-64x1.5B-Chat-Q4_K_M.gguf -ngl 99 -sm none -mg 0 -p "what is your name?" -n 30 -no-cnv

Problem description & steps to reproduce

root@alc-ai:/home/aubrey/work/llama-gpu# ./build/bin/llama-cli -m /srv/models/DeepSeek-V2-Lite-Chat-Q4_K_M/DeepSeek-V2-Lite-64x1.5B-Chat-Q4_K_M.gguf -ngl 99 -sm none -mg 0 -p "what is your name?" -n 30 -no-cnv
build: 4887 (8fcb563) with Intel(R) oneAPI DPC++/C++ Compiler 2025.0.4 (2025.0.4.20241205) for x86_64-unknown-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device SYCL0 (Intel(R) Arc(TM) A770 Graphics) - 15473 MiB free
llama_model_loader: loaded meta data with 47 key-value pairs and 377 tensors from /srv/models/DeepSeek-V2-Lite-Chat-Q4_K_M/DeepSeek-V2-Lite-64x1.5B-Chat-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = deepseek2
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = DeepSeek V2 Lite Chat
llama_model_loader: - kv 3: general.finetune str = Chat
llama_model_loader: - kv 4: general.basename str = DeepSeek-V2-Lite
llama_model_loader: - kv 5: general.size_label str = 64x1.5B
llama_model_loader: - kv 6: general.license str = other
llama_model_loader: - kv 7: general.license.name str = deepseek
llama_model_loader: - kv 8: general.license.link str = https://github.com/deepseek-ai/DeepSe...
llama_model_loader: - kv 9: deepseek2.block_count u32 = 27
llama_model_loader: - kv 10: deepseek2.context_length u32 = 163840
llama_model_loader: - kv 11: deepseek2.embedding_length u32 = 2048
llama_model_loader: - kv 12: deepseek2.feed_forward_length u32 = 10944
llama_model_loader: - kv 13: deepseek2.attention.head_count u32 = 16
llama_model_loader: - kv 14: deepseek2.attention.head_count_kv u32 = 16
llama_model_loader: - kv 15: deepseek2.rope.freq_base f32 = 10000.000000
llama_model_loader: - kv 16: deepseek2.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 17: deepseek2.expert_used_count u32 = 6
llama_model_loader: - kv 18: deepseek2.leading_dense_block_count u32 = 1
llama_model_loader: - kv 19: deepseek2.vocab_size u32 = 102400
llama_model_loader: - kv 20: deepseek2.attention.kv_lora_rank u32 = 512
llama_model_loader: - kv 21: deepseek2.attention.key_length u32 = 192
llama_model_loader: - kv 22: deepseek2.attention.value_length u32 = 128
llama_model_loader: - kv 23: deepseek2.expert_feed_forward_length u32 = 1408
llama_model_loader: - kv 24: deepseek2.expert_count u32 = 64
llama_model_loader: - kv 25: deepseek2.expert_shared_count u32 = 2
llama_model_loader: - kv 26: deepseek2.expert_weights_scale f32 = 1.000000
llama_model_loader: - kv 27: deepseek2.expert_weights_norm bool = false
llama_model_loader: - kv 28: deepseek2.expert_gating_func u32 = 1
llama_model_loader: - kv 29: deepseek2.rope.dimension_count u32 = 64
llama_model_loader: - kv 30: deepseek2.rope.scaling.type str = yarn
llama_model_loader: - kv 31: deepseek2.rope.scaling.factor f32 = 40.000000
llama_model_loader: - kv 32: deepseek2.rope.scaling.original_context_length u32 = 4096
llama_model_loader: - kv 33: deepseek2.rope.scaling.yarn_log_multiplier f32 = 0.070700
llama_model_loader: - kv 34: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 35: tokenizer.ggml.pre str = deepseek-llm
llama_model_loader: - kv 36: tokenizer.ggml.tokens arr[str,102400] = ["!", """, "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 37: tokenizer.ggml.token_type arr[i32,102400] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 38: tokenizer.ggml.merges arr[str,99757] = ["Ġ Ġ", "Ġ t", "Ġ a", "i n", "h e...
llama_model_loader: - kv 39: tokenizer.ggml.bos_token_id u32 = 100000
llama_model_loader: - kv 40: tokenizer.ggml.eos_token_id u32 = 100001
llama_model_loader: - kv 41: tokenizer.ggml.padding_token_id u32 = 100001
llama_model_loader: - kv 42: tokenizer.ggml.add_bos_token bool = true
llama_model_loader: - kv 43: tokenizer.ggml.add_eos_token bool = false
llama_model_loader: - kv 44: tokenizer.chat_template str = {% if not add_generation_prompt is de...
llama_model_loader: - kv 45: general.quantization_version u32 = 2
llama_model_loader: - kv 46: general.file_type u32 = 15
llama_model_loader: - type f32: 108 tensors
llama_model_loader: - type q5_0: 14 tensors
llama_model_loader: - type q8_0: 13 tensors
llama_model_loader: - type q4_K: 229 tensors
llama_model_loader: - type q6_K: 13 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q4_K - Medium
print_info: file size = 9.65 GiB (5.28 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 2
load: token to piece cache size = 0.6408 MB
print_info: arch = deepseek2
print_info: vocab_only = 0
print_info: n_ctx_train = 163840
print_info: n_embd = 2048
print_info: n_layer = 27
print_info: n_head = 16
print_info: n_head_kv = 16
print_info: n_rot = 64
print_info: n_swa = 0
print_info: n_swa_pattern = 1
print_info: n_embd_head_k = 192
print_info: n_embd_head_v = 128
print_info: n_gqa = 1
print_info: n_embd_k_gqa = 3072
print_info: n_embd_v_gqa = 2048
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-06
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 0.0e+00
print_info: n_ff = 10944
print_info: n_expert = 64
print_info: n_expert_used = 6
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 0
print_info: rope scaling = yarn
print_info: freq_base_train = 10000.0
print_info: freq_scale_train = 0.025
print_info: n_ctx_orig_yarn = 4096
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 0
print_info: ssm_d_inner = 0
print_info: ssm_d_state = 0
print_info: ssm_dt_rank = 0
print_info: ssm_dt_b_c_rms = 0
print_info: model type = 16B
print_info: model params = 15.71 B
print_info: general.name = DeepSeek V2 Lite Chat
print_info: n_layer_dense_lead = 1
print_info: n_lora_q = 0
print_info: n_lora_kv = 512
print_info: n_ff_exp = 1408
print_info: n_expert_shared = 2
print_info: expert_weights_scale = 1.0
print_info: expert_weights_norm = 0
print_info: expert_gating_func = softmax
print_info: rope_yarn_log_mul = 0.0707
print_info: vocab type = BPE
print_info: n_vocab = 102400
print_info: n_merges = 99757
print_info: BOS token = 100000 '<｜begin▁of▁sentence｜>'
print_info: EOS token = 100001 '<｜end▁of▁sentence｜>'
print_info: EOT token = 100001 '<｜end▁of▁sentence｜>'
print_info: PAD token = 100001 '<｜end▁of▁sentence｜>'
print_info: LF token = 185 'Ċ'
print_info: EOG token = 100001 '<｜end▁of▁sentence｜>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 27 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 28/28 layers to GPU
load_tensors: CPU_Mapped model buffer size = 112.50 MiB
load_tensors: SYCL0 model buffer size = 9767.98 MiB
.....................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = 0
llama_context: freq_base = 10000.0
llama_context: freq_scale = 0.025
llama_context: n_ctx_per_seq (4096) < n_ctx_train (163840) -- the full capacity of the model will not be utilized
Running with Environment Variables:
GGML_SYCL_DEBUG: 0
GGML_SYCL_DISABLE_OPT: 0
Build with Macros:
GGML_SYCL_FORCE_MMQ: no
GGML_SYCL_F16: no
Found 2 SYCL devices:
| | | | |Max | |Max |Global | |
| | | | |compute|Max work|sub |mem | |

ID	Device Type	Name	Version	units	group	group	size	Driver version
0	[level_zero:gpu:0]	Intel Arc A770 Graphics	12.55	512	1024	32	16225M	1.6.32224+14
1	[level_zero:gpu:1]	Intel UHD Graphics 770	12.2	32	512	32	62707M	1.6.32224+14
SYCL Optimization Feature:
ID	Device Type	Reorder
--	-------------------	-------
0	[level_zero:gpu:0]	Y
1	[level_zero:gpu:1]	N
llama_context: SYCL_Host output buffer size = 0.39 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 27, can_shift = 0
init: SYCL0 KV buffer size = 1080.00 MiB
llama_context: KV self size = 1080.00 MiB, K (f16): 648.00 MiB, V (f16): 432.00 MiB
llama_context: SYCL0 compute buffer size = 213.03 MiB
llama_context: SYCL_Host compute buffer size = 12.01 MiB
llama_context: graph nodes = 1924
llama_context: graph splits = 2
common_init_from_params: KV cache shifting is not supported for this context, disabling KV cache shifting
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 8

system_info: n_threads = 8 (n_threads_batch = 8) / 32 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 |

sampler seed: 2656463
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, n_batch = 2048, n_predict = 30, n_keep = 1

what is your name? is the difference between a man and a boy?
2 Answers | Add Yours
A man is an adult human male, while a boy is a

llama_perf_sampler_print: sampling time = 1.10 ms / 36 runs ( 0.03 ms per token, 32786.89 tokens per second)
llama_perf_context_print: load time = 3147.22 ms
llama_perf_context_print: prompt eval time = 288.22 ms / 6 tokens ( 48.04 ms per token, 20.82 tokens per second)
llama_perf_context_print: eval time = 1660.91 ms / 29 runs ( 57.27 ms per token, 17.46 tokens per second)
llama_perf_context_print: total time = 1952.91 ms / 35 tokens

First Bad Commit

No response

Relevant log output

[fairydreaming]

@aubreyli In non-interactive mode (-no-cnv) you have to include a proper prompt template in your prompt. For DeepSeek-V2 Lite it will be something like this: -p "User: what is your name?\n\nAssistant:"

[aubreyli]

@fairydreaming same issue here:

root@alc-ai:/home/aubrey/work/llama-gpu# ./build/bin/llama-cli -m /srv/models/DeepSeek-V2-Lite-Chat-Q4_K_M/DeepSeek-V2-Lite-64x1.5B-Chat-Q4_K_M.gguf -ngl 99 -sm none -mg 0 -p "User: what is your name?\n\nAssistant:" -n 30 -no-cnv
build: 4887 (8fcb563) with Intel(R) oneAPI DPC++/C++ Compiler 2025.0.4 (2025.0.4.20241205) for x86_64-unknown-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device SYCL0 (Intel(R) Arc(TM) A770 Graphics) - 15473 MiB free
llama_model_loader: loaded meta data with 47 key-value pairs and 377 tensors from /srv/models/DeepSeek-V2-Lite-Chat-Q4_K_M/DeepSeek-V2-Lite-64x1.5B-Chat-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.

--snip--

sampler seed: 457823384
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, n_batch = 2048, n_predict = 30, n_keep = 1

User: what is your name?

Assistant: This is my first time using this program. I have a question about how to use it.

Sure, I'd be happy to

llama_perf_sampler_print: sampling time = 1.37 ms / 42 runs ( 0.03 ms per token, 30746.71 tokens per second)
llama_perf_context_print: load time = 3257.36 ms
llama_perf_context_print: prompt eval time = 501.55 ms / 12 tokens ( 41.80 ms per token, 23.93 tokens per second)
llama_perf_context_print: eval time = 1659.25 ms / 29 runs ( 57.22 ms per token, 17.48 tokens per second)
llama_perf_context_print: total time = 2165.43 ms / 41 tokens

[fairydreaming]

@aubreyli Hmm, that's weird. Where can I download this model file?

[aubreyli]

@fairydreaming you can download it from here:
https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/tree/main
And convert it to Q4_K_M by the tools in llama.cpp.

[aubreyli]

@fairydreaming I have the same model files working properly by CUDA.

$ ./build/bin/llama-cli -m /srv/models/DeepSeek-V2-Lite-Chat-Q4_K_M/DeepSeek-V2-Lite-64x1.5B-Chat-Q4_K_M.gguf -ngl 99 -sm none -mg 0 -p "User: what is your name?\n\nAssistant:" -n 30 -no-cnv
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
build: 4790 (438a839) with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3060) - 11809 MiB free
llama_model_loader: loaded meta data with 47 key-value pairs and 377 tensors from /srv/models/DeepSeek-V2-Lite-Chat-Q4_K_M/DeepSeek-V2-Lite-64x1.5B-Chat-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.

----snip----

sampler seed: 563822659
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, n_batch = 2048, n_predict = 30, n_keep = 1

User: what is your name?

Assistant: I am DeepSeek Chat, an intelligent assistant developed by DeepSeek company. [end of text]

llama_perf_sampler_print: sampling time = 1.47 ms / 28 runs ( 0.05 ms per token, 19099.59 tokens per second)
llama_perf_context_print: load time = 2124.35 ms
llama_perf_context_print: prompt eval time = 81.80 ms / 12 tokens ( 6.82 ms per token, 146.69 tokens per second)
llama_perf_context_print: eval time = 197.45 ms / 15 runs ( 13.16 ms per token, 75.97 tokens per second)
llama_perf_context_print: total time = 289.15 ms / 27 tokens

[fairydreaming]

@aubreyli I confirm the problem, when using DeepSeek V2 Lite running on a GPU in a SYCL build (I used my RTX 4090 card) the model generates nonsense answers.

User: what is your name?

Assistant:entrenament d'equitació

Instruccions:

1. Llegeix l'escenari i escolta l'espectacle.

Model answers look somewhat coherent, but they seem to ignore the user prompt.

[qnixsynapse]

Please run it with GGML_SYCL_DEBUG=1 set in env.

[fairydreaming]

Please run it with GGML_SYCL_DEBUG=1 set in env.

sycl_debug.txt

[qnixsynapse]

Please run it with GGML_SYCL_DEBUG=1 set in env.

sycl_debug.txt

Interesting. Does test-backend-ops passes without fail?

[fairydreaming]

@qnixsynapse Yeah:

$ ./bin/test-backend-ops test -b SYCL0
Testing 2 devices

Backend 1/2: SYCL0
Running with Environment Variables:
  GGML_SYCL_DEBUG: 0
  GGML_SYCL_DISABLE_OPT: 0
Build with Macros:
  GGML_SYCL_FORCE_MMQ: no
  GGML_SYCL_F16: no
Found 1 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0|       [cuda:gpu:0]|                NVIDIA GeForce RTX 4090|    8.9|    128|    1024|   32| 25386M|            CUDA 12.4|
SYCL Optimization Feature:
|ID|        Device Type|Reorder|
|--|-------------------|-------|
| 0|       [cuda:gpu:0]|      N|
  Device description: NVIDIA GeForce RTX 4090
  Device memory: 24210 MB (23818 MB free)

  ABS(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  ABS(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  SGN(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  SGN(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  NEG(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  NEG(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  STEP(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  STEP(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  TANH(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  TANH(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  ELU(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  ELU(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  RELU(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  RELU(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  SIGMOID(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  SIGMOID(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  GELU(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  GELU(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  GELU_QUICK(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  GELU_QUICK(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  SILU(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  SILU(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  HARDSWISH(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  HARDSWISH(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  HARDSIGMOID(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  HARDSIGMOID(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  EXP(type=f16,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  EXP(type=f16,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  ABS(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  ABS(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  SGN(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  SGN(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  NEG(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  NEG(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  STEP(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  STEP(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  TANH(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  TANH(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  ELU(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  ELU(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  RELU(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  RELU(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  SIGMOID(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  SIGMOID(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  GELU(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  GELU(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  GELU_QUICK(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  GELU_QUICK(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  SILU(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  SILU(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  HARDSWISH(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  HARDSWISH(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  HARDSIGMOID(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  HARDSIGMOID(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  EXP(type=f16,ne_a=[128,2,2,2],v=1): not supported [SYCL0] 
  EXP(type=f16,ne_a=[5,7,11,13],v=1): not supported [SYCL0] 
  ABS(type=f32,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  ABS(type=f32,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  SGN(type=f32,ne_a=[128,2,2,2],v=0): not supported [SYCL0] 
  SGN(type=f32,ne_a=[5,7,11,13],v=0): not supported [SYCL0] 
  NEG(type=f32,ne_a=[128,2,2,2],v=0): OK
  NEG(type=f32,ne_a=[5,7,11,13],v=0): OK
  STEP(type=f32,ne_a=[128,2,2,2],v=0): OK
  STEP(type=f32,ne_a=[5,7,11,13],v=0): OK
  TANH(type=f32,ne_a=[128,2,2,2],v=0): OK
  TANH(type=f32,ne_a=[5,7,11,13],v=0): OK
...
  CROSS_ENTROPY_LOSS(type=f32,ne=[10,5,4,3]): not supported [SYCL0] 
  CROSS_ENTROPY_LOSS(type=f32,ne=[30000,1,1,1]): not supported [SYCL0] 
  CROSS_ENTROPY_LOSS_BACK(type=f32,ne=[10,5,4,3]): not supported [SYCL0] 
  CROSS_ENTROPY_LOSS_BACK(type=f32,ne=[30000,1,1,1]): not supported [SYCL0] 
  OPT_STEP_ADAMW(type=f32,ne=[10,5,4,3]): not supported [SYCL0] 
  4193/4193 tests passed
  Backend SYCL0: OK

In attached zip there are txt files with printed tensor values from CPU and SYCL backends. They seem to diverge more and more as the inference progresses.

tensors.zip

[fairydreaming]

I narrowed the problem to GGML_OP_MUL_MAT/GGML_OP_MUL_MAT_ID. When I remove them from the list of operations supported by SYCL everything starts working correctly.

[qnixsynapse]

They seem to diverge more and more as the inference progresses.

Does #12366 helps? I know there is a really bad memory leak in the backend.

Might also need to test with this: #12391

Unfortunately, I only have an A750 so can't test this model here.