Skip to content

Eval bug: Qwen 3.5 27B crashes running perplexity with RPC #19892

Description

@steampunque

Name and Version

llama.cpp b8148 on host and RPC machine

Operating systems

Linux

GGML backends

CUDA

Hardware

4070

Models

https://huggingface.co/steampunque/Qwen3.5-27B-MP-GGUF/blob/main/Qwen3.5-27B.Q4_K_H.gguf

Problem description & steps to reproduce

I found the model to intermittently crash when run with RPC on my downstream server. The problem can be replicated in upstream using a short perplexity run. The perplexity will run for awhile then eventually crash, same intermittent crashes I saw in my downstream server. My downstream server or perplexity will not crash if running locally without RPC.

llama-perplexity -m /data3hd/models/Qwen3.5-27B.Q4_K_H.gguf --rpc 192.9.200.4:50052 -ts .52028543291295045085,.47971456708704954914 -fit off --mmap -ngl 99 -c 1024 -b 128 -fa on --rope-scaling yarn --rope-scale 1.00000 --yarn-orig-ctx 262144 --rope_scale 1 -f short2.txt

short2.txt:
https://huggingface.co/steampunque/Qwen3.5-27B-MP-GGUF/blob/main/short2.txt

First Bad Commit

NA, new model

Relevant log output

Logs
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4070, compute capability 8.9, VMM: yes
DEPRECATED: argument '--rope-scale' specified multiple times, use comma-separated values instead (only last value will be used)
build: 8148 (244641955) with GNU 11.2.0 for Linux x86_64
llama_model_load_from_file_impl: using device RPC0 (192.9.200.4:50052) (unknown id) - 11844 MiB free
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4070) (0000:01:00.0) - 11558 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 851 tensors from /data3hd/models/Qwen3.5-27B.Q4_K_H.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              = qwen35
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 20
llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
llama_model_loader: - kv   4:                      general.sampling.temp f32              = 0.600000
llama_model_loader: - kv   5:                               general.name str              = Qwen3.5 27B
llama_model_loader: - kv   6:                           general.basename str              = Qwen3.5
llama_model_loader: - kv   7:                         general.size_label str              = 27B
llama_model_loader: - kv   8:                            general.license str              = apache-2.0
llama_model_loader: - kv   9:                       general.license.link str              = https://huggingface.co/Qwen/Qwen3.5-2...
llama_model_loader: - kv  10:                               general.tags arr[str,1]       = ["image-text-to-text"]
llama_model_loader: - kv  11:                         qwen35.block_count u32              = 64
llama_model_loader: - kv  12:                      qwen35.context_length u32              = 262144
llama_model_loader: - kv  13:                    qwen35.embedding_length u32              = 5120
llama_model_loader: - kv  14:                 qwen35.feed_forward_length u32              = 17408
llama_model_loader: - kv  15:                qwen35.attention.head_count u32              = 24
llama_model_loader: - kv  16:             qwen35.attention.head_count_kv u32              = 4
llama_model_loader: - kv  17:             qwen35.rope.dimension_sections arr[i32,4]       = [11, 11, 10, 0]
llama_model_loader: - kv  18:                      qwen35.rope.freq_base f32              = 10000000.000000
llama_model_loader: - kv  19:    qwen35.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  20:                qwen35.attention.key_length u32              = 256
llama_model_loader: - kv  21:              qwen35.attention.value_length u32              = 256
llama_model_loader: - kv  22:                     qwen35.ssm.conv_kernel u32              = 4
llama_model_loader: - kv  23:                      qwen35.ssm.state_size u32              = 128
llama_model_loader: - kv  24:                     qwen35.ssm.group_count u32              = 16
llama_model_loader: - kv  25:                  qwen35.ssm.time_step_rank u32              = 48
llama_model_loader: - kv  26:                      qwen35.ssm.inner_size u32              = 6144
llama_model_loader: - kv  27:             qwen35.full_attention_interval u32              = 4
llama_model_loader: - kv  28:                qwen35.rope.dimension_count u32              = 64
llama_model_loader: - kv  29:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  30:                         tokenizer.ggml.pre str              = qwen35
llama_model_loader: - kv  31:                      tokenizer.ggml.tokens arr[str,248320]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  32:                  tokenizer.ggml.token_type arr[i32,248320]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  33:                      tokenizer.ggml.merges arr[str,247587]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  34:                tokenizer.ggml.eos_token_id u32              = 248046
llama_model_loader: - kv  35:            tokenizer.ggml.padding_token_id u32              = 248044
llama_model_loader: - kv  36:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  37:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
llama_model_loader: - kv  38:               general.quantization_version u32              = 2
llama_model_loader: - kv  39:                          general.file_type u32              = 15
llama_model_loader: - type  f32:  353 tensors
llama_model_loader: - type q3_K:  118 tensors
llama_model_loader: - type q4_K:  261 tensors
llama_model_loader: - type q5_K:   79 tensors
llama_model_loader: - type q6_K:   40 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 14.85 GiB (4.74 BPW) 
load: 0 unused tokens
load: printing all EOG tokens:
load:   - 248044 ('<|endoftext|>')
load:   - 248046 ('<|im_end|>')
load:   - 248063 ('<|fim_pad|>')
load:   - 248064 ('<|repo_name|>')
load:   - 248065 ('<|file_sep|>')
load: special tokens cache size = 33
load: token to piece cache size = 1.7581 MB
print_info: arch                  = qwen35
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 5120
print_info: n_embd_inp            = 5120
print_info: n_layer               = 64
print_info: n_head                = 24
print_info: n_head_kv             = 4
print_info: n_rot                 = 64
print_info: n_swa                 = 0
print_info: is_swa_any            = 0
print_info: n_embd_head_k         = 256
print_info: n_embd_head_v         = 256
print_info: n_gqa                 = 6
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-06
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                  = 17408
print_info: n_expert              = 0
print_info: n_expert_used         = 0
print_info: n_expert_groups       = 0
print_info: n_group_used          = 0
print_info: causal attn           = 1
print_info: pooling type          = 0
print_info: rope type             = 40
print_info: rope scaling          = linear
print_info: freq_base_train       = 10000000.0
print_info: freq_scale_train      = 1
print_info: n_ctx_orig_yarn       = 262144
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: mrope sections        = [11, 11, 10, 0]
print_info: ssm_d_conv            = 4
print_info: ssm_d_inner           = 6144
print_info: ssm_d_state           = 128
print_info: ssm_dt_rank           = 48
print_info: ssm_n_group           = 16
print_info: ssm_dt_b_c_rms        = 0
print_info: model type            = ?B
print_info: model params          = 26.90 B
print_info: general.name          = Qwen3.5 27B
print_info: vocab type            = BPE
print_info: n_vocab               = 248320
print_info: n_merges              = 247587
print_info: BOS token             = 11 ','
print_info: EOS token             = 248046 '<|im_end|>'
print_info: EOT token             = 248046 '<|im_end|>'
print_info: PAD token             = 248044 '<|endoftext|>'
print_info: LF token              = 198 'Ċ'
print_info: FIM PRE token         = 248060 '<|fim_prefix|>'
print_info: FIM SUF token         = 248062 '<|fim_suffix|>'
print_info: FIM MID token         = 248061 '<|fim_middle|>'
print_info: FIM PAD token         = 248063 '<|fim_pad|>'
print_info: FIM REP token         = 248064 '<|repo_name|>'
print_info: FIM SEP token         = 248065 '<|file_sep|>'
print_info: EOG token             = 248044 '<|endoftext|>'
print_info: EOG token             = 248046 '<|im_end|>'
print_info: EOG token             = 248063 '<|fim_pad|>'
print_info: EOG token             = 248064 '<|repo_name|>'
print_info: EOG token             = 248065 '<|file_sep|>'
print_info: max token length      = 256
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 63 repeating layers to GPU
load_tensors: offloaded 65/65 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   682.03 MiB
load_tensors:        CUDA0 model buffer size =  7827.38 MiB
load_tensors: RPC0[192.9.200.4:50052] model buffer size =  6700.44 MiB
...........................................................................................
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 1024
llama_context: n_ctx_seq     = 1024
llama_context: n_batch       = 128
llama_context: n_ubatch      = 128
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = false
llama_context: freq_base     = 10000000.0
llama_context: freq_scale    = 1
llama_context: yarn_log_mul  = 0
llama_context: n_ctx_seq (1024) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.95 MiB
llama_kv_cache:      CUDA0 KV buffer size =    32.00 MiB
llama_kv_cache: RPC0[192.9.200.4:50052] KV buffer size =    32.00 MiB
llama_kv_cache: size =   64.00 MiB (  1024 cells,  16 layers,  1/1 seqs), K (f16):   32.00 MiB, V (f16):   32.00 MiB
llama_memory_recurrent:      CUDA0 RS buffer size =    68.58 MiB
llama_memory_recurrent: RPC0[192.9.200.4:50052] RS buffer size =    81.05 MiB
llama_memory_recurrent: size =  149.62 MiB (     1 cells,  64 layers,  1 seqs), R (f32):    5.62 MiB, S (f32):  144.00 MiB
sched_reserve: reserving ...
sched_reserve: RPC0[192.9.200.4:50052] compute buffer size =    42.12 MiB
sched_reserve:      CUDA0 compute buffer size =   123.75 MiB
sched_reserve:        CPU compute buffer size =     5.50 MiB
sched_reserve: graph nodes  = 8409 (with bs=128), 4713 (with bs=1)
sched_reserve: graph splits = 3
sched_reserve: reserve took 134.68 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 800,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
perplexity: tokenizing the input ..
perplexity: tokenization took 149.382 ms
perplexity: calculating perplexity over 62 chunks, n_ctx=1024, batch_size=128, n_seq=1
perplexity: 2.29 seconds per pass - ETA 2.35 minutes
[1]4.1892,[2]5.7518,[3]6.3636,[4]6.7647,[5]7.3672,[6]7.4310,[7]7.2039,[8]7.4688,[9]7.5061,[10]7.2696,[11]7.1021,[12]6.4889,[13]6.1404,[14]5.9566,[15]5.7864,[16]5.6947,[17]5.7200,[18]5.8061,[19]5.8059,[20]5.9134,[21]5.9285,[22]5.9513,[23]5.8930,[24]6.0645,[25]6.1147,[26]6.1385,[27]6.2020,[28]6.2878,[29]6.2774,[30]6.2977,[31]6.3160,[32]6.3730,[33]6.4608,[34]6.4459,[35]6.4121,[36]6.4383,[37]6.4734,[38]6.4661,[39]6.4770,[40]6.3189,[41]6.3403,[42]6.3267,[43]6.3997,[44]6.3826,[45]6.3455,[46]6.3259,[47]6.3400,[48]6.3222,[49]6.3246,[50]6.3910,[51]6.3983,[52]6.3786,[53]6.3971,/usr/local/src/ai/lm/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp:669: Remote RPC server crashed or returned malformed response
recv failed (bytes_recv=0, size_to_recv=8)
[New LWP 14848]
[New LWP 14853]
[New LWP 14854]
[New LWP 14855]

[short2.txt](https://github.com/user-attachments/files/25547358/short2.txt)

[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".
0x00007fde7d1a83c7 in wait4 () from /lib64/libc.so.6
#0  0x00007fde7d1a83c7 in wait4 () from /lib64/libc.so.6
#1  0x00007fde7d641a0b in ggml_print_backtrace () from /usr/lib64/libggml-base.so.0
#2  0x00007fde7d641b5d in ggml_abort () from /usr/lib64/libggml-base.so.0
#3  0x00007fde7d6d79a2 in ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer*, ggml_tensor const*, void*, unsigned long, unsigned long) () from /usr/lib64/libggml-rpc.so.0
#4  0x00007fde7d6581d6 in ggml_backend_tensor_copy () from /usr/lib64/libggml-base.so.0
#5  0x00007fde7d65d026 in ggml_backend_sched_graph_compute_async () from /usr/lib64/libggml-base.so.0
#6  0x00007fde84c1e259 in llama_context::graph_compute(ggml_cgraph*, bool) () from /usr/lib64/libllama.so.0
#7  0x00007fde84c1e635 in llama_context::process_ubatch(llama_ubatch const&, llm_graph_type, llama_memory_context_i*, ggml_status&) () from /usr/lib64/libllama.so.0
#8  0x00007fde84c25e03 in llama_context::decode(llama_batch const&) () from /usr/lib64/libllama.so.0
#9  0x00007fde84c273fc in llama_decode () from /usr/lib64/libllama.so.0
#10 0x0000000000474004 in perplexity(llama_context*, common_params const&, int) ()
#11 0x000000000046bd1e in main ()
[Inferior 1 (process 14847) detached]
ll.start: line 1: 14847 Aborted                 llama-perplexity -m /data3hd/models/Qwen3.5-27B.Q4_K_H.gguf --rpc 192.9.200.4:50052 -ts .52028543291295045085,.47971456708704954914 -fit off --mmap -ngl 99 -c 1024 -b 128 -fa on --rope-scaling yarn --rope-scale 1.00000 --yarn-orig-ctx 262144 --rope_scale 1 -f short2.txt

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions