AMD 7600XT - llama off loads to GPU but low TpS

[frankmtl-git]

I've been trying to use an AMD 7600XT with Pi5.
I got it up and running with the self-compiled kernel and AMD drivers.

DISPLAY=:0 glxinfo -B gives:

name of display: :0
display: :0  screen: 0
direct rendering: Yes
Extended renderer info (GLX_MESA_query_renderer):
    Vendor: AMD (0x1002)
    Device: AMD Radeon RX 7600 XT (gfx1102, LLVM 15.0.6, DRM 3.54, 6.6.70-v8-16k+) (0x7480)
    Version: 23.2.1
    Accelerated: yes
    Video memory: 16384MB
    Unified memory: no
    Preferred profile: core (0x1)
    Max core profile version: 4.6
    Max compat profile version: 4.6
    Max GLES1 profile version: 1.1
    Max GLES[23] profile version: 3.2
Memory info (GL_ATI_meminfo):
    VBO free memory - total: 16321 MB, largest block: 16321 MB
    VBO free aux. memory - total: 3958 MB, largest block: 3958 MB
    Texture free memory - total: 16321 MB, largest block: 16321 MB
    Texture free aux. memory - total: 3958 MB, largest block: 3958 MB
    Renderbuffer free memory - total: 16321 MB, largest block: 16321 MB
    Renderbuffer free aux. memory - total: 3958 MB, largest block: 3958 MB
Memory info (GL_NVX_gpu_memory_info):
    Dedicated video memory: 16384 MB
    Total available memory: 20410 MB
    Currently available dedicated video memory: 16321 MB
OpenGL vendor string: AMD
OpenGL renderer string: AMD Radeon RX 7600 XT (gfx1102, LLVM 15.0.6, DRM 3.54, 6.6.70-v8-16k+)
OpenGL core profile version string: 4.6 (Core Profile) Mesa 23.2.1-1~bpo12+rpt3
OpenGL core profile shading language version string: 4.60
OpenGL core profile context flags: (none)
OpenGL core profile profile mask: core profile

OpenGL version string: 4.6 (Compatibility Profile) Mesa 23.2.1-1~bpo12+rpt3
OpenGL shading language version string: 4.60
OpenGL context flags: (none)
OpenGL profile mask: compatibility profile

OpenGL ES profile version string: OpenGL ES 3.2 Mesa 23.2.1-1~bpo12+rpt3
OpenGL ES profile shading language version string: OpenGL ES GLSL ES 3.20``

When I launch ``./build/bin/llama-cli -m models/Llama-3.2-3B-Instruct-Q4_K_M.gguf -n 50 -e -ngl 33 -t 4

I get:

ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 7600 XT (RADV GFX1102) (radv) | uma: 0 | fp16: 1 | warp size: 64 | matrix cores: none
build: 4563 (caf773f2) with cc (Debian 12.2.0-14) 12.2.0 for aarch64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon RX 7600 XT (RADV GFX1102)) - 16128 MiB free
llama_model_loader: loaded meta data with 35 key-value pairs and 255 tensors from models/Llama-3.2-3B-Instruct-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              = llama
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Llama 3.2 3B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Llama-3.2
llama_model_loader: - kv   5:                         general.size_label str              = 3B
llama_model_loader: - kv   6:                            general.license str              = llama3.2
llama_model_loader: - kv   7:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv   8:                          general.languages arr[str,8]       = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv   9:                          llama.block_count u32              = 28
llama_model_loader: - kv  10:                       llama.context_length u32              = 131072
llama_model_loader: - kv  11:                     llama.embedding_length u32              = 3072
llama_model_loader: - kv  12:                  llama.feed_forward_length u32              = 8192
llama_model_loader: - kv  13:                 llama.attention.head_count u32              = 24
llama_model_loader: - kv  14:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  15:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  16:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  17:                 llama.attention.key_length u32              = 128
llama_model_loader: - kv  18:               llama.attention.value_length u32              = 128
llama_model_loader: - kv  19:                          general.file_type u32              = 15
llama_model_loader: - kv  20:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  21:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  22:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  23:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  24:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  26:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  27:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  28:                tokenizer.ggml.eos_token_id u32              = 128009
llama_model_loader: - kv  29:                    tokenizer.chat_template str              = {{- bos_token }}\n{%- if custom_tools ...
llama_model_loader: - kv  30:               general.quantization_version u32              = 2
llama_model_loader: - kv  31:                      quantize.imatrix.file str              = /models_out/Llama-3.2-3B-Instruct-GGU...
llama_model_loader: - kv  32:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  33:             quantize.imatrix.entries_count i32              = 196
llama_model_loader: - kv  34:              quantize.imatrix.chunks_count i32              = 125
llama_model_loader: - type  f32:   58 tensors
llama_model_loader: - type q4_K:  168 tensors
llama_model_loader: - type q6_K:   29 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 1.87 GiB (5.01 BPW)
load: special tokens cache size = 256
load: token to piece cache size = 0.7999 MB
print_info: arch             = llama
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 3072
print_info: n_layer          = 28
print_info: n_head           = 24
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 3
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-05
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: n_ff             = 8192
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = linear
print_info: freq_base_train  = 500000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 131072
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       = 3B
print_info: model params     = 3.21 B
print_info: general.name     = Llama 3.2 3B Instruct
print_info: vocab type       = BPE
print_info: n_vocab          = 128256
print_info: n_merges         = 280147
print_info: BOS token        = 128000 '<|begin_of_text|>'
print_info: EOS token        = 128009 '<|eot_id|>'
print_info: EOT token        = 128009 '<|eot_id|>'
print_info: EOM token        = 128008 '<|eom_id|>'
print_info: LF token         = 128 'Ä'
print_info: EOG token        = 128008 '<|eom_id|>'
print_info: EOG token        = 128009 '<|eot_id|>'
print_info: max token length = 256
load_tensors: offloading 28 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 29/29 layers to GPU
load_tensors:      Vulkan0 model buffer size =  1918.35 MiB
load_tensors:   CPU_Mapped model buffer size =   308.23 MiB
llama_init_from_model: n_seq_max     = 1
llama_init_from_model: n_ctx         = 4096
llama_init_from_model: n_ctx_per_seq = 4096
llama_init_from_model: n_batch       = 2048
llama_init_from_model: n_ubatch      = 512
llama_init_from_model: flash_attn    = 0
llama_init_from_model: freq_base     = 500000.0
llama_init_from_model: freq_scale    = 1
llama_init_from_model: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 28, can_shift = 1
llama_kv_cache_init:    Vulkan0 KV buffer size =   448.00 MiB
llama_init_from_model: KV self size  =  448.00 MiB, K (f16):  224.00 MiB, V (f16):  224.00 MiB
llama_init_from_model: Vulkan_Host  output buffer size =     0.49 MiB
llama_init_from_model:    Vulkan0 compute buffer size =   256.50 MiB
llama_init_from_model: Vulkan_Host compute buffer size =    14.01 MiB
llama_init_from_model: graph nodes  = 902
llama_init_from_model: graph splits = 2
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)

but performance are really low and i can't really use the LLM:

llama_perf_sampler_print:    sampling time =       1.01 ms /    14 runs   (    0.07 ms per token, 13806.71 tokens per second)
llama_perf_context_print:        load time =   36552.88 ms
llama_perf_context_print: prompt eval time =   27448.31 ms /    21 tokens ( 1307.06 ms per token,     0.77 tokens per second)
llama_perf_context_print:        eval time =    7666.34 ms /     3 runs   ( 2555.45 ms per token,     0.39 tokens per second)
llama_perf_context_print:       total time =   38405.26 ms /    24 tokens

If I also open nvtop everything goes smoothly and i can have normal conversation but i still get:

llama_perf_sampler_print:    sampling time =       6.37 ms /    21 runs   (    0.30 ms per token,  3297.74 tokens per second)
llama_perf_context_print:        load time =   40915.43 ms
llama_perf_context_print: prompt eval time =   78729.57 ms /    91 tokens (  865.16 ms per token,     1.16 tokens per second)
llama_perf_context_print:        eval time =   12532.62 ms /   156 runs   (   80.34 ms per token,    12.45 tokens per second)
llama_perf_context_print:       total time =  230074.09 ms /   247 tokens

which seems to me a bit odd since conversation is really fluid.

Also if I try a beefier model (mistral-7b-instruct-v0.2.Q8_0.gguf), I can't even load it on to the GPU even though the model is just around 7GB:

radv/amdgpu: Failed to allocate a buffer:
radv/amdgpu:    size      : 4248715264 bytes
radv/amdgpu:    alignment : 262144 bytes
radv/amdgpu:    domains   : 4
ggml_vulkan: Device memory allocation of size 4248715264 failed.
ggml_vulkan: vk::Device::allocateMemory: ErrorOutOfDeviceMemory
llama_model_load: error loading model: unable to allocate Vulkan0 buffer
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model 'models/mistral-7b-instruct-v0.2.Q8_0.gguf'
main: error: unable to load model

not sure if this is just vulkan fault though.

[geerlingguy]

You may need to adjust the GGML_VK_FORCE_MAX_ALLOCATION_SIZE size due to some bugs in the driver on arm64. See: geerlingguy/ollama-benchmark#1 and have a read through my notes towards the bottom of that original issue post.

[frankmtl-git]

Thanks! export GGML_VK_FORCE_MAX_ALLOCATION_SIZE=2147483647 made me load the 7B model correctly.
I still don't understand why i need nvtop open to make any model use the GPU but at least it work