ggml_cuda_compute_forward: RMS_NORM failed

[AnonymousVibrate]

llm = LlamaCpp(
    model_path=model_name_or_path,
    n_ctx= 2048,
    verbose=True,
    n_threads=4,
    n_batch=512,
    n_gpu_layers = 8,
    callback_manager=callback_manager,
    stop = ['HUMAN:'], # Dynamic stopping when such token is detected.
    temperature = 0.4,
    streaming=True
)

This might be because of n_batch =512
Output : ggml_cuda_compute_forward: RMS_NORM failed

llm = LlamaCpp(
    model_path=model_name_or_path,
    n_ctx= 2048,
    verbose=True,
    n_threads=4,
    n_batch=30,
    n_gpu_layers = 8,
    callback_manager=callback_manager,
    stop = ['HUMAN:'], # Dynamic stopping when such token is detected.
    temperature = 0.4,
    streaming=True
)

Output : ggml_cuda_compute_forward: ADD failed

TERMINAL :

---

llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.name str              = .
llama_model_loader: - kv   2:                       llama.context_length u32              = 2048
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 3200
llama_model_loader: - kv   4:                          llama.block_count u32              = 26
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 8640
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 100
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 32
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  10:                          general.file_type u32              = 2
llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  12:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  13:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  15:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  17:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  18:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   53 tensors
llama_model_loader: - type q4_0:  183 tensors
llama_model_loader: - type q8_0:    1 tensors
llm_load_vocab: special tokens definition check successful ( 259/32000 ).
llm_load_print_meta: format           = GGUF V2
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 2048
llm_load_print_meta: n_embd           = 3200
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_layer          = 26
llm_load_print_meta: n_rot            = 100
llm_load_print_meta: n_embd_head_k    = 100
llm_load_print_meta: n_embd_head_v    = 100
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 3200
llm_load_print_meta: n_embd_v_gqa     = 3200
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-06
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: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 8640
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 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: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: model type       = 3B
llm_load_print_meta: model ftype      = Q4_0
llm_load_print_meta: model params     = 3.43 B
llm_load_print_meta: model size       = 1.84 GiB (4.62 BPW)
llm_load_print_meta: general.name     = .
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce 930MX, compute capability 5.0, VMM: yes
llm_load_tensors: ggml ctx size =    0.18 MiB
llm_load_tensors: offloading 8 repeating layers to GPU
llm_load_tensors: offloaded 8/27 layers to GPU
llm_load_tensors:        CPU buffer size =  1887.49 MiB
llm_load_tensors:      CUDA0 buffer size =   531.95 MiB
..............................................................................................
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: n_batch    = 30
llama_new_context_with_model: n_ubatch   = 30
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:  CUDA_Host KV buffer size =   450.00 MiB
llama_kv_cache_init:      CUDA0 KV buffer size =   200.00 MiB
llama_new_context_with_model: KV self size  =  650.00 MiB, K (f16):  325.00 MiB, V (f16):  325.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.12 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =     9.20 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =     9.20 MiB
llama_new_context_with_model: graph nodes  = 838
llama_new_context_with_model: graph splits = 3
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 | MATMUL_INT8 = 0 |
Model metadata: {'general.name': '.', 'general.architecture': 'llama', 'llama.context_length': '2048', 'llama.rope.dimension_count': '100', 'llama.embedding_length': '3200', 'llama.block_count': '26', 'llama.feed_forward_length': '8640', 'llama.attention.head_count': '32', 'tokenizer.ggml.eos_token_id': '2', 'general.file_type': '2', 'llama.attention.head_count_kv': '32', 'llama.attention.layer_norm_rms_epsilon': '0.000001', 'tokenizer.ggml.model': 'llama', 'general.quantization_version': '2', 'tokenizer.ggml.bos_token_id': '1', 'tokenizer.ggml.padding_token_id': '0'}
Using fallback chat format: None

---

There is no issue if not offloaded to GPU. It runs smoothly on my CPU but it takes time to generate.
Im using Model : orca-mini-3b-gguf2-q4_0.gguf (version GGUF V2)

[abetlen]

@AnonymousVibrate do you mind testing this with the llama.cpp cli as well, this may be an upstream issue.

[AnonymousVibrate]

@AnonymousVibrate do you mind testing this with the llama.cpp cli as well, this may be an upstream issue.

Thank you ! I fixed it by upgrading my GPU drivers to latest

[amaiya]

If anyone stumbles on this, I received similar errors like the following using WSL2 on Windows 11 (depending on the model being loaded):

ggml_cuda_compute_forward: RMS_NORM failed
CUDA error: the provided PTX was compiled with an unsupported toolchain.
/tmp/pip-install-8m0m2p29/llama-cpp-python_a7006c57169d4ac7b422996f2f70b0ac/vendor/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:70: CUDA error

The solution was to install llama-cpp-python as follows:

CUDACXX=/usr/local/cuda-12/bin/nvcc CMAKE_ARGS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=all-major" FORCE_CMAKE=1 pip install llama-cpp-python --no-cache-dir --force-reinstall --upgrade