Eval bug: phi-4 crashes with new versions

[porzione]

Name and Version

llama-cli --version
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
register_backend: registered backend CUDA (1 devices)
register_device: registered device CUDA0 (NVIDIA GeForce RTX 3060)
register_backend: registered backend BLAS (1 devices)
register_device: registered device BLAS (OpenBLAS)
register_backend: registered backend RPC (0 devices)
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (AMD Ryzen 7 3800X 8-Core Processor)
version: 5435 (a4090d117)
built with cc (GCC) 15.1.1 20250425 for x86_64-pc-linux-gnu

Operating systems

Linux

GGML backends

CUDA

Hardware

Nvidia 3060 12Gb
AMD Ryzen 7 3800X 8-Core

Models

Original MS build phi-4-Q4_K.gguf

Problem description & steps to reproduce

Last working version is b5426, after that llama-cli crashes with 'Segmentation fault'

Command: llama-cli -m /home/ftp/AI/microsoft/phi-4-gguf/phi-4-Q4_K.gguf -st --simple-io --flash-attn --no-display-prompt -ngl 41 --threads 8 --temp 0.25 --top-p 0.95 --ctx-size 2048 -p "$(cat ~/dox/AI/PROMPT.txt)"

Workaround: downgrade to b5426

First Bad Commit

b5429 (supposedly)

Relevant log output

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
register_backend: registered backend CUDA (1 devices)
register_device: registered device CUDA0 (NVIDIA GeForce RTX 3060)
register_backend: registered backend BLAS (1 devices)
register_device: registered device BLAS (OpenBLAS)
register_backend: registered backend RPC (0 devices)
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (AMD Ryzen 7 3800X 8-Core Processor)
build: 5435 (a4090d117) with cc (GCC) 15.1.1 20250425 for x86_64-pc-linux-gnu (debug)
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) - 10769 MiB free
llama_model_loader: loaded meta data with 33 key-value pairs and 243 tensors from /home/ftp/AI/microsoft/phi-4-gguf/phi-4-Q4_K.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              = phi3
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Phi 4
llama_model_loader: - kv   3:                            general.version str              = 4
llama_model_loader: - kv   4:                       general.organization str              = Microsoft
llama_model_loader: - kv   5:                           general.basename str              = phi
llama_model_loader: - kv   6:                         general.size_label str              = 15B
llama_model_loader: - kv   7:                            general.license str              = mit
llama_model_loader: - kv   8:                       general.license.link str              = https://huggingface.co/microsoft/phi-...
llama_model_loader: - kv   9:                               general.tags arr[str,7]       = ["phi", "nlp", "math", "code", "chat"...
llama_model_loader: - kv  10:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  11:                        phi3.context_length u32              = 16384
llama_model_loader: - kv  12:  phi3.rope.scaling.original_context_length u32              = 16384
llama_model_loader: - kv  13:                      phi3.embedding_length u32              = 5120
llama_model_loader: - kv  14:                   phi3.feed_forward_length u32              = 17920
llama_model_loader: - kv  15:                           phi3.block_count u32              = 40
llama_model_loader: - kv  16:                  phi3.attention.head_count u32              = 40
llama_model_loader: - kv  17:               phi3.attention.head_count_kv u32              = 10
llama_model_loader: - kv  18:      phi3.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  19:                  phi3.rope.dimension_count u32              = 128
llama_model_loader: - kv  20:                        phi3.rope.freq_base f32              = 250000.000000
llama_model_loader: - kv  21:              phi3.attention.sliding_window u32              = 0
llama_model_loader: - kv  22:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  23:                         tokenizer.ggml.pre str              = dbrx
llama_model_loader: - kv  24:                      tokenizer.ggml.tokens arr[str,100352]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,100352]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  26:                      tokenizer.ggml.merges arr[str,100000]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  27:                tokenizer.ggml.bos_token_id u32              = 100257
llama_model_loader: - kv  28:                tokenizer.ggml.eos_token_id u32              = 100265
llama_model_loader: - kv  29:            tokenizer.ggml.padding_token_id u32              = 100349
llama_model_loader: - kv  30:                    tokenizer.chat_template str              = {% for message in messages %}{% if (m...
llama_model_loader: - kv  31:               general.quantization_version u32              = 2
llama_model_loader: - kv  32:                          general.file_type u32              = 15
llama_model_loader: - type  f32:   81 tensors
llama_model_loader: - type q4_K:  101 tensors
llama_model_loader: - type q5_K:   40 tensors
llama_model_loader: - type q6_K:   21 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 8.43 GiB (4.94 BPW) 
load: special tokens cache size = 96
load: token to piece cache size = 0.6151 MB
print_info: arch             = phi3
print_info: vocab_only       = 0
print_info: n_ctx_train      = 16384
print_info: n_embd           = 5120
print_info: n_layer          = 40
print_info: n_head           = 40
print_info: n_head_kv        = 10
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 4
print_info: n_embd_k_gqa     = 1280
print_info: n_embd_v_gqa     = 1280
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: f_attn_scale     = 0.0e+00
print_info: n_ff             = 17920
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        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 250000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 16384
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       = 14B
print_info: model params     = 14.66 B
print_info: general.name     = Phi 4
print_info: vocab type       = BPE
print_info: n_vocab          = 100352
print_info: n_merges         = 100000
print_info: BOS token        = 100257 '<|endoftext|>'
print_info: EOS token        = 100265 '<|im_end|>'
print_info: EOT token        = 100265 '<|im_end|>'
print_info: PAD token        = 100349 '<|dummy_85|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 100258 '<|fim_prefix|>'
print_info: FIM SUF token    = 100260 '<|fim_suffix|>'
print_info: FIM MID token    = 100259 '<|fim_middle|>'
print_info: EOG token        = 100257 '<|endoftext|>'
print_info: EOG token        = 100265 '<|im_end|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 40 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 41/41 layers to GPU
load_tensors:        CUDA0 model buffer size =  8354.71 MiB
load_tensors:   CPU_Mapped model buffer size =   275.62 MiB
.......................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 2048
llama_context: n_ctx_per_seq = 2048
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 250000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (2048) < n_ctx_train (16384) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.38 MiB
llama_kv_cache_unified:      CUDA0 KV buffer size =   400.00 MiB
llama_kv_cache_unified: size =  400.00 MiB (  2048 cells,  40 layers), K (f16):  200.00 MiB, V (f16):  200.00 MiB
Segmentation fault (core dumped)

[savchenko]

Can confirm, but the traceback is slightly different. With ROCm:

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)
/home/user/code/llama.cpp/src/llama-kv-cache.cpp:673: GGML_ASSERT(swa_type != LLAMA_SWA_TYPE_NONE && "do not prune non-SWA cache") failed
[New LWP 406709]
[New LWP 406708]
[New LWP 406707]
[New LWP 406706]
[New LWP 406705]
[New LWP 406704]
[New LWP 406703]
[New LWP 406702]
[New LWP 406701]
[New LWP 406700]
[New LWP 406699]
[New LWP 406667]
[New LWP 406666]
[New LWP 406665]
[New LWP 406664]
[New LWP 406663]
[New LWP 406662]
[New LWP 406661]
[New LWP 406660]
[New LWP 406659]
[New LWP 406658]
[New LWP 406657]
[New LWP 406656]
[New LWP 406655]
[New LWP 406654]
[New LWP 406653]
[New LWP 406652]
[New LWP 406651]
[New LWP 406650]
[New LWP 406649]
[New LWP 406648]
[New LWP 406647]
[New LWP 406646]
[New LWP 406645]
[New LWP 406644]
[New LWP 406643]
[New LWP 406639]
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
__syscall_cancel_arch () at ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S:56
warning: 56	../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S: No such file or directory
#0  __syscall_cancel_arch () at ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S:56
56	in ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S
#1  0x00007f0b13099668 in __internal_syscall_cancel (a1=<optimized out>, a2=<optimized out>, a3=<optimized out>, a4=<optimized out>, a5=a5@entry=0, a6=a6@entry=0, nr=61) at ./nptl/cancellation.c:49
warning: 49	./nptl/cancellation.c: No such file or directory
#2  0x00007f0b130996ad in __syscall_cancel (a1=<optimized out>, a2=<optimized out>, a3=<optimized out>, a4=<optimized out>, a5=a5@entry=0, a6=a6@entry=0, nr=61) at ./nptl/cancellation.c:75
75	in ./nptl/cancellation.c
#3  0x00007f0b13104787 in __GI___wait4 (pid=<optimized out>, stat_loc=<optimized out>, options=<optimized out>, usage=<optimized out>) at ../sysdeps/unix/sysv/linux/wait4.c:30
warning: 30	../sysdeps/unix/sysv/linux/wait4.c: No such file or directory
#4  0x00007f0b13600eb8 in ggml_abort () from /home/user/code/llama.cpp/build/bin/libggml-base.so
#5  0x00007f0b1376c247 in llama_kv_cache_unified::prune_swa(int, int, int) () from /home/user/code/llama.cpp/build/bin/libllama.so
#6  0x00007f0b1376c3f3 in llama_kv_cache_unified_iswa::commit() () from /home/user/code/llama.cpp/build/bin/libllama.so
#7  0x00007f0b1373d267 in llama_context::decode(llama_batch&) () from /home/user/code/llama.cpp/build/bin/libllama.so
#8  0x00007f0b1373e623 in llama_decode () from /home/user/code/llama.cpp/build/bin/libllama.so
#9  0x000055f4553f2da4 in common_init_from_params(common_params&) ()
#10 0x000055f4552aa11e in server_context::load_model(common_params const&) ()
#11 0x000055f45524edfc in main ()
[Inferior 1 (process 406638) detached]
Aborted (core dumped)

[porzione]

yes, amd vulkan crashes too
I've tried another, older build 'bartowski/phi-4-Q4_K_M.gguf' - the same

[lcarrere]

Same here,

@ggerganov

the phi 4 mini is OK. The phi 14B is initializing kv-cache as llama_kv_cache_unified instance.

later it is accessing it through :
const auto * kv = is_swa ? kv_self->get_kv_swa() : kv_self->get_kv_base(); //is_swa is false here

But kv_self->get_kv_base() returns an untitialized area and the first subequent access throws a segfault here:

const int32_t ikv = map_layer_ids.at(il);

[ggerganov]

Can you confirm that #13676 fixes the issues?

[lcarrere]

Can you confirm that #13676 fixes the issues?

No crash anymore with it though I have some doubts about the quality of completions. I'm compiling in release and running a benchmark. This will take 10 minutes

[lcarrere]

@ggerganov all good!