Phi-2 q4_km generating gibberish on ARM devices

[LostRuins]

Running the latest commit, testing the model https://huggingface.co/afrideva/phi-2-uncensored-GGUF/blob/main/phi-2-uncensored.q4_k_m.gguf in Termux.

./main -m ../phi-2-uncensored.q4_k_m.gguf -n 10 -p "Hi, my name is"

main: build = 1691 (7082d24)
main: built with clang version 17.0.6 for aarch64-unknown-linux-android24
main: seed  = 1703404966
llama_model_loader: loaded meta data with 20 key-value pairs and 325 tensors from ../phi-2-uncensored.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              = 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              = 15
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 q4_K:   81 tensors
llama_model_loader: - type q5_K:   32 tensors
llama_model_loader: - type q6_K:   17 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_gqa            = 1
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      = Q4_K - Medium
llm_load_print_meta: model params     = 2.78 B
llm_load_print_meta: model size       = 1.66 GiB (5.14 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         = 128 '�'
llm_load_tensors: ggml ctx size       =    0.12 MiB
llm_load_tensors: system memory used  = 1704.75 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_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 = 4 / 8 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | 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 = 10, n_keep = 0

Hi, my name is vocwich Reeves TABLEeco Feahar Reeves sill Reeves

It generates complete gibberish.

The same model works fine on my x86_64 windows device.
Also, q2_k works fine on both systems.

Is it possible that that intrinsic ARM FMA or ARM NEON is responsible for this issue?

Also tagging @ebeyabraham

[ggerganov]

Not sure - works on M2 Ultra

What is the output from:

make tests && ./tests/test-backend-ops

Any fails?

[LostRuins]

Sorry, that took really long to compile on my Android.

The tests crash halfway, consistently at the same point, right after MOE(n_experts=8,n_experts_per_tok=2,n_tokens=1,n_embd=4096,n_ff=8192) .
I am not sure if it's due to going OOM, a segfault or something else. Here's the terminal output:

tests.txt

I modified the cpp file to remove https://github.com/ggerganov/llama.cpp/blob/master/tests/test-backend-ops.cpp#L1580 and recompiled.

tests_2.txt

And all other tests seem to be passing for test-backend-ops. I think the MOE test is unrelated though, Phi is not a MOE model. Also to clarify, the main binary is made from the makefile without modifications, just running make main. The environment is Termux on Android 9.

[LostRuins]

Is there an easy way to modify the makefile to skip either NEON or FMA? I'd like to see if I can pinpoint which one is causing issues.

Also, I did check with another user over discord, they used Termux with identical compile and run settings to me and their output was coherent. So it might be a device specific thing? Slightly confused and wondering if anybody else has the same issue.

For reference, my device (that didn't work)

SoC: Samsung Exynos 9810 (S5E9810)
CPU: 4x 2.9 GHz Exynos M3 Meerkat, 4x 1.9 GHz ARM Cortex-A55, Cores: 8

[ggerganov]

Does it work with this patch:

diff --git a/ggml-quants.c b/ggml-quants.c
index a15a2404..b5c76f00 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -5602,7 +5602,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     const int nb = n / QK_K;
 
-#ifdef __ARM_NEON
+#ifdef __ARM_NEON_XXX
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
 

[LostRuins]

Unfortunately not, it's still generating rubbish.

[LostRuins]

Another update: I have done a full search and replace in all files replacing all instances of __ARM_NEON to __ARM_NEON_XXX and it is working correctly now. So whatever the bug was, it was indeed with NEON.

I will try to slowly replace each instance until I find the one responsible, unless you have a better approach to suggest.

[LostRuins]

By trial and error, I have narrowed it down to ggml_vec_dot_q5_K_q8_K()

By skipping https://github.com/ggerganov/llama.cpp/blob/master/ggml-quants.c#L5879 , this model works perfectly on my device.

[ggerganov]

Could you verify that #4630 also fixes the issue on that device?

[LostRuins]

Yes that seems to have fixed it! Awesome.

Though I am wondering why the unit tests didn't catch that.

[lts-rad]

I checked q4_k_m on my build with #4630 and thats fine, but I think there's something still off with q5_k_m. https://huggingface.co/afrideva/phi-2-uncensored-GGUF/blob/main/phi-2-uncensored.q5_k_m.gguf

[ggerganov]

Though I am wondering why the unit tests didn't catch that.

The tests cannot catch integer overflows. Should be fixed now