Eval bug: Running llama-server only possible with single AMD GPU, running multiple always causes Segmentation fault regardless of model size

[Zubkover]

Name and Version

Core tl;dr of the issue:

• Having single GPU = 20b 4bit OSS loads and runs with no issues
• Running 2/3/4 = no model suddenly is able to be be loaded and all attempts regardless of llama-server params end up in Segmentation fault (core dumped) error

Side issues/questions:

• Do you explicitly need to have all GPUs installed at build time? I noticed that having single working GPU (on which I built initial llama.cpp on) and then installing 3 more causes llama-server to auto-crash with immediate effect and Killed console output (due to ROCm failing with

ERROR Received I2C_NAK_7B0ADDR_NOACK
ERROR WriteI2CData() - I2C error occured
Failed to read EEPROM table header

Honestly speaking, don't know what I'm doing wrong and I'm 3 days deep into trying to understand based on docs and tutorials if I'm missing something in build/llama-server params 🙈

What I tried:

• played around with llama-server params, including of/off for no-mmap and ngl
• played around with build commands, including solution from Eval bug: usr/lib/gcc/x86_64-pc-linux-gnu/14/include/g++-v14/bits/random.tcc:2668: void std::discrete_distribution<_IntType>::param_type::_M_initialize() [with _IntType = int]: Assertion '__sum > 0' failed. #15551
• updated motherboard BIOS from P1.10 to P1.50
• switched BIOSes Radeon VII <> Radeon Pro VII
• reinstalled Ubuntu 4 times and repeated setup
• tried tensorfiles from rocblas 6.4.4 and 7.1.0-2 (both while running ROCm 7.1.1 as dedicated ROCm 5.6 cannot be installed on Ubuntu 24 apparently)

Operating systems

Linux

GGML backends

BLAS

Hardware

• Ubuntu 24.04.3 LTS (fresh and clean install, only for llama.cpp purposes)
• AsRock H510 PRO BTC+ (@ P1.50 BIOS)
• Celeron G5905
• 4GB RAM
• 4x Radeon Pro VII (gfx906)

Models

No response

Problem description & steps to reproduce

1. Install more than 1 AMD GPU
2. Run any model, even one that would fit within 1 GPU's memory
3. llama-server ends with Segmentation fault (core dumped)

First Bad Commit

No response

Relevant log output

llama-server -m /home/llm/llama.cpp/models/gpt-oss-20b-Q8_0.gguf --host 0.0.0.0 --port 8080

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 4 ROCm devices:
  Device 0: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 1: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 2: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 3: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
main: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)
build: 7179 (4abef75f2) with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu
system info: n_threads = 2, n_threads_batch = 2, total_threads = 2

system_info: n_threads = 2 (n_threads_batch = 2) / 2 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

init: using 6 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/llm/llama.cpp/models/gpt-oss-20b-Q8_0.gguf'
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon (TM) Pro VII) (0000:03:00.0) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm1 (AMD Radeon (TM) Pro VII) (0000:06:00.0) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm2 (AMD Radeon (TM) Pro VII) (0000:09:00.0) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm3 (AMD Radeon (TM) Pro VII) (0000:0c:00.0) - 16348 MiB free
llama_model_loader: loaded meta data with 37 key-value pairs and 459 tensors from /home/llm/llama.cpp/models/gpt-oss-20b-Q8_0.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              = gpt-oss
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Gpt-Oss-20B
llama_model_loader: - kv   3:                           general.basename str              = Gpt-Oss-20B
llama_model_loader: - kv   4:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   5:                         general.size_label str              = 20B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   8:                               general.tags arr[str,2]       = ["vllm", "text-generation"]
llama_model_loader: - kv   9:                        gpt-oss.block_count u32              = 24
llama_model_loader: - kv  10:                     gpt-oss.context_length u32              = 131072
llama_model_loader: - kv  11:                   gpt-oss.embedding_length u32              = 2880
llama_model_loader: - kv  12:                gpt-oss.feed_forward_length u32              = 2880
llama_model_loader: - kv  13:               gpt-oss.attention.head_count u32              = 64
llama_model_loader: - kv  14:            gpt-oss.attention.head_count_kv u32              = 8
llama_model_loader: - kv  15:                     gpt-oss.rope.freq_base f32              = 150000.000000
llama_model_loader: - kv  16:   gpt-oss.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  17:                       gpt-oss.expert_count u32              = 32
llama_model_loader: - kv  18:                  gpt-oss.expert_used_count u32              = 4
llama_model_loader: - kv  19:               gpt-oss.attention.key_length u32              = 64
llama_model_loader: - kv  20:             gpt-oss.attention.value_length u32              = 64
llama_model_loader: - kv  21:           gpt-oss.attention.sliding_window u32              = 128
llama_model_loader: - kv  22:         gpt-oss.expert_feed_forward_length u32              = 2880
llama_model_loader: - kv  23:                  gpt-oss.rope.scaling.type str              = yarn
llama_model_loader: - kv  24:                gpt-oss.rope.scaling.factor f32              = 32.000000
llama_model_loader: - kv  25: gpt-oss.rope.scaling.original_context_length u32              = 4096
llama_model_loader: - kv  26:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  27:                         tokenizer.ggml.pre str              = gpt-4o
llama_model_loader: - kv  28:                      tokenizer.ggml.tokens arr[str,201088]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  29:                  tokenizer.ggml.token_type arr[i32,201088]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  30:                      tokenizer.ggml.merges arr[str,446189]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  31:                tokenizer.ggml.bos_token_id u32              = 199998
llama_model_loader: - kv  32:                tokenizer.ggml.eos_token_id u32              = 200002
llama_model_loader: - kv  33:            tokenizer.ggml.padding_token_id u32              = 200017
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {# Chat template fixes by Unsloth #}\n...
llama_model_loader: - kv  35:               general.quantization_version u32              = 2
llama_model_loader: - kv  36:                          general.file_type u32              = 7
llama_model_loader: - type  f32:  289 tensors
llama_model_loader: - type q8_0:   98 tensors
llama_model_loader: - type mxfp4:   72 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q8_0
print_info: file size   = 11.27 GiB (4.63 BPW) 
load: printing all EOG tokens:
load:   - 199999 ('<|endoftext|>')
load:   - 200002 ('<|return|>')
load:   - 200007 ('<|end|>')
load:   - 200012 ('<|call|>')
load: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list
load: special tokens cache size = 21
load: token to piece cache size = 1.3332 MB
print_info: arch             = gpt-oss
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 2880
print_info: n_embd_inp       = 2880
print_info: n_layer          = 24
print_info: n_head           = 64
print_info: n_head_kv        = 8
print_info: n_rot            = 64
print_info: n_swa            = 128
print_info: is_swa_any       = 1
print_info: n_embd_head_k    = 64
print_info: n_embd_head_v    = 64
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 512
print_info: n_embd_v_gqa     = 512
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             = 2880
print_info: n_expert         = 32
print_info: n_expert_used    = 4
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        = 2
print_info: rope scaling     = yarn
print_info: freq_base_train  = 150000.0
print_info: freq_scale_train = 0.03125
print_info: n_ctx_orig_yarn  = 4096
print_info: rope_finetuned   = unknown
print_info: model type       = 20B
print_info: model params     = 20.91 B
print_info: general.name     = Gpt-Oss-20B
print_info: n_ff_exp         = 2880
print_info: vocab type       = BPE
print_info: n_vocab          = 201088
print_info: n_merges         = 446189
print_info: BOS token        = 199998 '<|startoftext|>'
print_info: EOS token        = 200002 '<|return|>'
print_info: EOT token        = 199999 '<|endoftext|>'
print_info: PAD token        = 200017 '<|reserved_200017|>'
print_info: LF token         = 198 'Ċ'
print_info: EOG token        = 199999 '<|endoftext|>'
print_info: EOG token        = 200002 '<|return|>'
print_info: EOG token        = 200012 '<|call|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 24 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 25/25 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   586.82 MiB
load_tensors:        ROCm0 model buffer size =  3022.41 MiB
load_tensors:        ROCm1 model buffer size =  2590.64 MiB
load_tensors:        ROCm2 model buffer size =  2590.64 MiB
load_tensors:        ROCm3 model buffer size =  2745.70 MiB
........................Segmentation fault (core dumped)

[pwilkin]

Please recompile with DEBUG mode and add a stracktrace for the segmentation fault.

[Zubkover]

cmake -B build -DCMAKE_BUILD_TYPE=Debug I assume? 👀

If so then output is as follows:

llama-server -m /home/llm/llama.cpp/models/gpt-oss-20b-Q8_0.gguf --host 0.0.0.0 --port 8080
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 4 ROCm devices:
  Device 0: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 1: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 2: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 3: AMD Radeon (TM) Pro VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
register_backend: registered backend ROCm (4 devices)
register_device: registered device ROCm0 (AMD Radeon (TM) Pro VII)
register_device: registered device ROCm1 (AMD Radeon (TM) Pro VII)
register_device: registered device ROCm2 (AMD Radeon (TM) Pro VII)
register_device: registered device ROCm3 (AMD Radeon (TM) Pro VII)
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (Intel(R) Celeron(R) G5905 CPU @ 3.50GHz)
load_backend: failed to find ggml_backend_init in /home/llm/llama.cpp/build/bin/libggml-hip.so
load_backend: failed to find ggml_backend_init in /home/llm/llama.cpp/build/bin/libggml-cpu.so
main: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)
build: 7179 (4abef75f2) with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu (debug)
system info: n_threads = 2, n_threads_batch = 2, total_threads = 2

system_info: n_threads = 2 (n_threads_batch = 2) / 2 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

init: using 6 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/llm/llama.cpp/models/gpt-oss-20b-Q8_0.gguf'
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon (TM) Pro VII) (0000:03:00.0) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm1 (AMD Radeon (TM) Pro VII) (0000:06:00.0) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm2 (AMD Radeon (TM) Pro VII) (0000:09:00.0) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm3 (AMD Radeon (TM) Pro VII) (0000:0c:00.0) - 16348 MiB free
llama_model_loader: loaded meta data with 37 key-value pairs and 459 tensors from /home/llm/llama.cpp/models/gpt-oss-20b-Q8_0.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              = gpt-oss
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Gpt-Oss-20B
llama_model_loader: - kv   3:                           general.basename str              = Gpt-Oss-20B
llama_model_loader: - kv   4:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   5:                         general.size_label str              = 20B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   8:                               general.tags arr[str,2]       = ["vllm", "text-generation"]
llama_model_loader: - kv   9:                        gpt-oss.block_count u32              = 24
llama_model_loader: - kv  10:                     gpt-oss.context_length u32              = 131072
llama_model_loader: - kv  11:                   gpt-oss.embedding_length u32              = 2880
llama_model_loader: - kv  12:                gpt-oss.feed_forward_length u32              = 2880
llama_model_loader: - kv  13:               gpt-oss.attention.head_count u32              = 64
llama_model_loader: - kv  14:            gpt-oss.attention.head_count_kv u32              = 8
llama_model_loader: - kv  15:                     gpt-oss.rope.freq_base f32              = 150000.000000
llama_model_loader: - kv  16:   gpt-oss.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  17:                       gpt-oss.expert_count u32              = 32
llama_model_loader: - kv  18:                  gpt-oss.expert_used_count u32              = 4
llama_model_loader: - kv  19:               gpt-oss.attention.key_length u32              = 64
llama_model_loader: - kv  20:             gpt-oss.attention.value_length u32              = 64
llama_model_loader: - kv  21:           gpt-oss.attention.sliding_window u32              = 128
llama_model_loader: - kv  22:         gpt-oss.expert_feed_forward_length u32              = 2880
llama_model_loader: - kv  23:                  gpt-oss.rope.scaling.type str              = yarn
llama_model_loader: - kv  24:                gpt-oss.rope.scaling.factor f32              = 32.000000
llama_model_loader: - kv  25: gpt-oss.rope.scaling.original_context_length u32              = 4096
llama_model_loader: - kv  26:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  27:                         tokenizer.ggml.pre str              = gpt-4o
llama_model_loader: - kv  28:                      tokenizer.ggml.tokens arr[str,201088]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  29:                  tokenizer.ggml.token_type arr[i32,201088]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  30:                      tokenizer.ggml.merges arr[str,446189]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  31:                tokenizer.ggml.bos_token_id u32              = 199998
llama_model_loader: - kv  32:                tokenizer.ggml.eos_token_id u32              = 200002
llama_model_loader: - kv  33:            tokenizer.ggml.padding_token_id u32              = 200017
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {# Chat template fixes by Unsloth #}\n...
llama_model_loader: - kv  35:               general.quantization_version u32              = 2
llama_model_loader: - kv  36:                          general.file_type u32              = 7
llama_model_loader: - type  f32:  289 tensors
llama_model_loader: - type q8_0:   98 tensors
llama_model_loader: - type mxfp4:   72 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q8_0
print_info: file size   = 11.27 GiB (4.63 BPW) 
load: printing all EOG tokens:
load:   - 199999 ('<|endoftext|>')
load:   - 200002 ('<|return|>')
load:   - 200007 ('<|end|>')
load:   - 200012 ('<|call|>')
load: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list
load: special tokens cache size = 21
load: token to piece cache size = 1.3332 MB
print_info: arch             = gpt-oss
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 2880
print_info: n_embd_inp       = 2880
print_info: n_layer          = 24
print_info: n_head           = 64
print_info: n_head_kv        = 8
print_info: n_rot            = 64
print_info: n_swa            = 128
print_info: is_swa_any       = 1
print_info: n_embd_head_k    = 64
print_info: n_embd_head_v    = 64
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 512
print_info: n_embd_v_gqa     = 512
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             = 2880
print_info: n_expert         = 32
print_info: n_expert_used    = 4
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        = 2
print_info: rope scaling     = yarn
print_info: freq_base_train  = 150000.0
print_info: freq_scale_train = 0.03125
print_info: n_ctx_orig_yarn  = 4096
print_info: rope_finetuned   = unknown
print_info: model type       = 20B
print_info: model params     = 20.91 B
print_info: general.name     = Gpt-Oss-20B
print_info: n_ff_exp         = 2880
print_info: vocab type       = BPE
print_info: n_vocab          = 201088
print_info: n_merges         = 446189
print_info: BOS token        = 199998 '<|startoftext|>'
print_info: EOS token        = 200002 '<|return|>'
print_info: EOT token        = 199999 '<|endoftext|>'
print_info: PAD token        = 200017 '<|reserved_200017|>'
print_info: LF token         = 198 'Ċ'
print_info: EOG token        = 199999 '<|endoftext|>'
print_info: EOG token        = 200002 '<|return|>'
print_info: EOG token        = 200012 '<|call|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 24 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 25/25 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   586.82 MiB
load_tensors:        ROCm0 model buffer size =  3022.41 MiB
load_tensors:        ROCm1 model buffer size =  2590.64 MiB
load_tensors:        ROCm2 model buffer size =  2590.64 MiB
load_tensors:        ROCm3 model buffer size =  2745.70 MiB
........................Segmentation fault (core dumped)

and dump, but it still has??despite being from debug build?

Core was generated by `llama-server -m /home/llm/llama.cpp/models/gpt-oss-20b-Q8_0.gguf --host 0.0.0.0'.
Program terminated with signal SIGSEGV, Segmentation fault.
#0  0x000073c4cb44e424 in ?? ()
[Current thread is 1 (LWP 10604)]
(gdb) bt
#0  0x000073c4cb44e424 in ?? ()
#1  0x00007ffe5d470690 in ?? ()
#2  0x000073c5042addae in ?? ()
#3  0x00006173b2390600 in ?? ()
#4  0x28be8ff8ab9e7800 in ?? ()
#5  0x00007ffe5d4707e0 in ?? ()
#6  0x00006173abd1ec10 in ?? ()
#7  0x00007ffe5d4707e0 in ?? ()
#8  0x0000000000000000 in ?? ()

[JohannesGaessler]

I am unable to reproduce the issue using a MI50 + an RX 6800.