CLBlast segfaulting with Phi q3_k_m and GPU offload (broken after backend integration)

[LostRuins]

CLBlast is segfaulting when using phi-2-uncensored.q3_k_m.gguf in the latest release, since the backend integration.

C:\test> main.exe -ngl 8 -m E:\LLaMA\models\phi-2-uncensored.q3_k_m.gguf -p "Hi, my name is"
Log start
main: build = 1961 (1387ea21)
main: built with MSVC 19.37.32826.1 for x64
main: seed  = 1706113455
ggml_opencl: selecting platform: 'NVIDIA CUDA'
ggml_opencl: selecting device: 'NVIDIA GeForce RTX 2060'
ggml_opencl: device FP16 support: false
llama_model_loader: loaded meta data with 20 key-value pairs and 325 tensors from E:\LLaMA\models\phi-2-uncensored.q3_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              = phi2
llama_model_loader: - kv   1:                               general.name str              = Phi2
llama_model_loader: - kv   2:                        phi2.context_length u32              = 2048
llama_model_loader: - kv   3:                      phi2.embedding_length u32              = 2560
llama_model_loader: - kv   4:                   phi2.feed_forward_length u32              = 10240
llama_model_loader: - kv   5:                           phi2.block_count u32              = 32
llama_model_loader: - kv   6:                  phi2.attention.head_count u32              = 32
llama_model_loader: - kv   7:               phi2.attention.head_count_kv u32              = 32
llama_model_loader: - kv   8:          phi2.attention.layer_norm_epsilon f32              = 0.000010
llama_model_loader: - kv   9:                  phi2.rope.dimension_count u32              = 32
llama_model_loader: - kv  10:                          general.file_type u32              = 12
llama_model_loader: - kv  11:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,51200]   = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,51200]   = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  15:                      tokenizer.ggml.merges arr[str,50000]   = ["Ġ t", "Ġ a", "h e", "i n", "r e",...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 50256
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 50256
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 50256
llama_model_loader: - kv  19:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  195 tensors
llama_model_loader: - type q3_K:   33 tensors
llama_model_loader: - type q4_K:   94 tensors
llama_model_loader: - type q5_K:    2 tensors
llama_model_loader: - type q6_K:    1 tensors
llm_load_vocab: mismatch in special tokens definition ( 910/51200 vs 944/51200 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = phi2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 51200
llm_load_print_meta: n_merges         = 50000
llm_load_print_meta: n_ctx_train      = 2048
llm_load_print_meta: n_embd           = 2560
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 32
llm_load_print_meta: n_embd_head_k    = 80
llm_load_print_meta: n_embd_head_v    = 80
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 2560
llm_load_print_meta: n_embd_v_gqa     = 2560
llm_load_print_meta: f_norm_eps       = 1.0e-05
llm_load_print_meta: f_norm_rms_eps   = 0.0e+00
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: n_ff             = 10240
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 2048
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: model type       = 3B
llm_load_print_meta: model ftype      = Q3_K - Medium
llm_load_print_meta: model params     = 2.78 B
llm_load_print_meta: model size       = 1.38 GiB (4.25 BPW)
llm_load_print_meta: general.name     = Phi2
llm_load_print_meta: BOS token        = 50256 '<|endoftext|>'
llm_load_print_meta: EOS token        = 50256 '<|endoftext|>'
llm_load_print_meta: UNK token        = 50256 '<|endoftext|>'
llm_load_print_meta: LF token         = 30 '?'
llm_load_tensors: ggml ctx size =    0.25 MiB
llm_load_tensors: offloading 8 repeating layers to GPU
llm_load_tensors: offloaded 8/33 layers to GPU
llm_load_tensors:        CPU buffer size =  1317.88 MiB
llm_load_tensors:     OpenCL buffer size =   311.88 MiB
............................................................................................
llama_new_context_with_model: n_ctx      = 512
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =   160.00 MiB
llama_new_context_with_model: KV self size  =  160.00 MiB, K (f16):   80.00 MiB, V (f16):   80.00 MiB
llama_new_context_with_model:        CPU input buffer size   =     6.01 MiB
llama_new_context_with_model:        CPU compute buffer size =   105.00 MiB
llama_new_context_with_model: graph splits (measure): 1

the program segfaults at this point

I've tested build = 1844 (3fe8178) and it fails too.

It seems to be broken since the backend integration. This model loaded successfully previously before build 1842. Currently, it still loads fine if 0 layers are offloaded, otherwise it segfaults. Other llama based models seem to work fine.

Earlier successful build:

C:\test2> main.exe -ngl 8 -m E:\LLaMA\models\phi-2-uncensored.q3_k_m.gguf -p "Hi, my name is"
Log start
main: build = 1842 (584d674)
main: built with MSVC 19.37.32826.1 for x64
main: seed  = 1706114088
ggml_opencl: selecting platform: 'NVIDIA CUDA'
ggml_opencl: selecting device: 'NVIDIA GeForce RTX 2060'
ggml_opencl: device FP16 support: false
llama_model_loader: loaded meta data with 20 key-value pairs and 325 tensors from E:\LLaMA\models\phi-2-uncensored.q3_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              = phi2
llama_model_loader: - kv   1:                               general.name str              = Phi2
llama_model_loader: - kv   2:                        phi2.context_length u32              = 2048
llama_model_loader: - kv   3:                      phi2.embedding_length u32              = 2560
llama_model_loader: - kv   4:                   phi2.feed_forward_length u32              = 10240
llama_model_loader: - kv   5:                           phi2.block_count u32              = 32
llama_model_loader: - kv   6:                  phi2.attention.head_count u32              = 32
llama_model_loader: - kv   7:               phi2.attention.head_count_kv u32              = 32
llama_model_loader: - kv   8:          phi2.attention.layer_norm_epsilon f32              = 0.000010
llama_model_loader: - kv   9:                  phi2.rope.dimension_count u32              = 32
llama_model_loader: - kv  10:                          general.file_type u32              = 12
llama_model_loader: - kv  11:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,51200]   = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,51200]   = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  15:                      tokenizer.ggml.merges arr[str,50000]   = ["Ġ t", "Ġ a", "h e", "i n", "r e",...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 50256
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 50256
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 50256
llama_model_loader: - kv  19:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  195 tensors
llama_model_loader: - type q3_K:   33 tensors
llama_model_loader: - type q4_K:   94 tensors
llama_model_loader: - type q5_K:    2 tensors
llama_model_loader: - type q6_K:    1 tensors
llm_load_vocab: mismatch in special tokens definition ( 910/51200 vs 944/51200 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = phi2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 51200
llm_load_print_meta: n_merges         = 50000
llm_load_print_meta: n_ctx_train      = 2048
llm_load_print_meta: n_embd           = 2560
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 32
llm_load_print_meta: n_embd_head_k    = 80
llm_load_print_meta: n_embd_head_v    = 80
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 2560
llm_load_print_meta: n_embd_v_gqa     = 2560
llm_load_print_meta: f_norm_eps       = 1.0e-05
llm_load_print_meta: f_norm_rms_eps   = 0.0e+00
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: n_ff             = 10240
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 2048
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: model type       = 3B
llm_load_print_meta: model ftype      = Q3_K - Medium
llm_load_print_meta: model params     = 2.78 B
llm_load_print_meta: model size       = 1.38 GiB (4.25 BPW)
llm_load_print_meta: general.name     = Phi2
llm_load_print_meta: BOS token        = 50256 '<|endoftext|>'
llm_load_print_meta: EOS token        = 50256 '<|endoftext|>'
llm_load_print_meta: UNK token        = 50256 '<|endoftext|>'
llm_load_print_meta: LF token         = 30 '?'
llm_load_tensors: ggml ctx size       =    0.12 MiB
llm_load_tensors: using OpenCL for GPU acceleration
llm_load_tensors: system memory used  = 1098.23 MiB
llm_load_tensors: VRAM used           =  311.80 MiB
llm_load_tensors: offloading 8 repeating layers to GPU
llm_load_tensors: offloaded 8/33 layers to GPU
............................................................................................
llama_new_context_with_model: n_ctx      = 512
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_new_context_with_model: KV self size  =  160.00 MiB, K (f16):   80.00 MiB, V (f16):   80.00 MiB
llama_build_graph: non-view tensors processed: 774/774
llama_new_context_with_model: compute buffer total size = 113.19 MiB

system_info: n_threads = 6 / 12 | AVX = 1 | AVX_VNNI = 0 | 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 = 0 | VSX = 0 |
sampling:
        repeat_last_n = 64, repeat_penalty = 1.100, frequency_penalty = 0.000, presence_penalty = 0.000
        top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order:
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temp
generate: n_ctx = 512, n_batch = 512, n_predict = -1, n_keep = 0


Hi, my name is John and I'm a software engineer. I work for a small company that develops custom software

[sorasoras]

sadly, Opencl has always been broken. it's just that before backend integration, CPU act as backup with mmap.

#2059 (comment)

[LostRuins]

I'm not sure that's applicable, I don't think Phi is a MoE model anyway.

[slaren]

It's the same issue as with MoE. Phi uses bias tensors, but OpenCL doesn't support the ggml_add operation.

[LostRuins]

@sorasoras so just wanna double-check with you, after the backend changes, Phi, Mixtral, Falcon, all not working with OpenCL anymore correct?

@slaren in the meantime, any ideas on how to detect if a model would use an unsupported operation before loading it? For Phi, I can check the arch and warn the user, but for Mixtral, the reported arch is simply reported 'llama' which would give no indication it's unsupported up till it segfaults and crashes (and it won't be clear why either).

you earlier mentioned in #2059

...so that the weights not supported by a backend are kept on the CPU instead that would somewhat restore this behavior

if you can point me to where in code I might be able to test for this, would be awesome.

Also tagging @0cc4m if you might have any ideas.

[slaren]

if you can point me to where in code I might be able to test for this, would be awesome.

The relevant code is here:
https://github.com/ggerganov/llama.cpp/blob/5f1925a8cef81eb9b372faaae34b0dd76d5361d4/llama.cpp#L3414-L3418

A simple solution would be to use a different set of contexts for the bias and MoE tensors. The ml.create_tensor calls for MoE and bias tensors would need to be changed to use these contexts instead. When using OpenCL, this context would be set to the CPU buffer. I really would prefer to not do this to avoid complicating the code too much, but if that's something that you are interested in doing in kobold.cpp, that would probably be the easiest solution.

[LostRuins]

Closing as resolved after Occam added ggml_add

Edit: was already closed