Qwen3-Next --ubatch-size issue

[2432896620-ctrl]

I find another problem.:
When I adjust the size of the parameter -ub (--ubatch-size) to more than 512, I get the memory error:

ggml_new_object: not enough space in the context's memory pool (needed 10711552, available 10711184)

Hardware and Software Environment

• GPU: AMD Radeon Graphics
• GPU Architecture: gfx1151 (Device ID: 0x1151)
• Driver/Stack: ROCm (reported as NO_VMM = 1)
• CPU: x86_64 with AVX512 support
• OS: Linux (Ubuntu)
• llama.cpp Build Info: Built with gcc 15.2.0, ROCm backend enabled.

mark@MarkPC:~/llama.cpp/llama.cpp-master$ ./llama-server -m /home/mark/Models/Q8/Qwen3-Next-80B-A3B-Instruct-Q8_0/Qwen3-Next-80B-A3B-Instruct-Q8_0.gguf -fa 1 -c 65536 --host 0.0.0.0 --port 8090 -ub 4096 --no-mmap
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
main: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)
build: 0 (unknown) with cc (Ubuntu 15.2.0-4ubuntu4) 15.2.0 for x86_64-linux-gnu
system info: n_threads = 16, n_threads_batch = 16, total_threads = 32

system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |

init: using 31 threads for HTTP server
start: binding port with default address family
main: loading model
srv load_model: loading model '/home/mark/Models/Q8/Qwen3-Next-80B-A3B-Instruct-Q8_0/Qwen3-Next-80B-A3B-Instruct-Q8_0.gguf'
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon Graphics) (0000:c6:00.0) - 121499 MiB free
llama_model_loader: loaded meta data with 44 key-value pairs and 807 tensors from /home/mark/Models/Q8/Qwen3-Next-80B-A3B-Instruct-Q8_0/Qwen3-Next-80B-A3B-Instruct-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 = qwen3next
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 = 0.800000
llama_model_loader: - kv 4: general.sampling.temp f32 = 0.700000
llama_model_loader: - kv 5: general.name str = Qwen3 Next A3B Instruct
llama_model_loader: - kv 6: general.finetune str = Instruct
llama_model_loader: - kv 7: general.basename str = Qwen3-Next
llama_model_loader: - kv 8: general.size_label str = A3B
llama_model_loader: - kv 9: general.license str = apache-2.0
llama_model_loader: - kv 10: general.license.link str = https://huggingface.co/Qwen/Qwen3-Nex...
llama_model_loader: - kv 11: general.tags arr[str,1] = ["text-generation"]
llama_model_loader: - kv 12: qwen3next.block_count u32 = 48
llama_model_loader: - kv 13: qwen3next.context_length u32 = 262144
llama_model_loader: - kv 14: qwen3next.embedding_length u32 = 2048
llama_model_loader: - kv 15: qwen3next.feed_forward_length u32 = 5120
llama_model_loader: - kv 16: qwen3next.attention.head_count u32 = 16
llama_model_loader: - kv 17: qwen3next.attention.head_count_kv u32 = 2
llama_model_loader: - kv 18: qwen3next.rope.freq_base f32 = 10000000.000000
llama_model_loader: - kv 19: qwen3next.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 20: qwen3next.expert_used_count u32 = 10
llama_model_loader: - kv 21: qwen3next.attention.key_length u32 = 256
llama_model_loader: - kv 22: qwen3next.attention.value_length u32 = 256
llama_model_loader: - kv 23: general.file_type u32 = 7
llama_model_loader: - kv 24: qwen3next.expert_count u32 = 512
llama_model_loader: - kv 25: qwen3next.expert_feed_forward_length u32 = 512
llama_model_loader: - kv 26: qwen3next.expert_shared_feed_forward_length u32 = 512
llama_model_loader: - kv 27: qwen3next.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 28: qwen3next.ssm.state_size u32 = 128
llama_model_loader: - kv 29: qwen3next.ssm.group_count u32 = 16
llama_model_loader: - kv 30: qwen3next.ssm.time_step_rank u32 = 32
llama_model_loader: - kv 31: qwen3next.ssm.inner_size u32 = 4096
llama_model_loader: - kv 32: qwen3next.rope.dimension_count u32 = 64
llama_model_loader: - kv 33: general.quantization_version u32 = 2
llama_model_loader: - kv 34: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 35: tokenizer.ggml.pre str = qwen2
llama_model_loader: - kv 36: tokenizer.ggml.tokens arr[str,151936] = ["!", """, "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 37: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 38: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 39: tokenizer.ggml.eos_token_id u32 = 151645
llama_model_loader: - kv 40: tokenizer.ggml.padding_token_id u32 = 151643
llama_model_loader: - kv 41: tokenizer.ggml.bos_token_id u32 = 151643
llama_model_loader: - kv 42: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 43: tokenizer.chat_template str = {%- if tools %}\n {{-'<|im_start|>...
llama_model_loader: - type f32: 313 tensors
llama_model_loader: - type q8_0: 494 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q8_0
print_info: file size = 78.98 GiB (8.52 BPW)
load: printing all EOG tokens:
load: - 151643 ('<|endoftext|>')
load: - 151645 ('<|im_end|>')
load: - 151662 ('<|fim_pad|>')
load: - 151663 ('<|repo_name|>')
load: - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch = qwen3next
print_info: vocab_only = 0
print_info: n_ctx_train = 262144
print_info: n_embd = 2048
print_info: n_embd_inp = 2048
print_info: n_layer = 48
print_info: n_head = 16
print_info: n_head_kv = 2
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 = 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-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 = 5120
print_info: n_expert = 512
print_info: n_expert_used = 10
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 = linear
print_info: freq_base_train = 10000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 262144
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 4
print_info: ssm_d_inner = 4096
print_info: ssm_d_state = 128
print_info: ssm_dt_rank = 32
print_info: ssm_n_group = 16
print_info: ssm_dt_b_c_rms = 0
print_info: model type = ?B
print_info: model params = 79.67 B
print_info: general.name = Qwen3 Next A3B Instruct
print_info: vocab type = BPE
print_info: n_vocab = 151936
print_info: n_merges = 151387
print_info: BOS token = 151643 '<|endoftext|>'
print_info: EOS token = 151645 '<|im_end|>'
print_info: EOT token = 151645 '<|im_end|>'
print_info: PAD token = 151643 '<|endoftext|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 151659 '<|fim_prefix|>'
print_info: FIM SUF token = 151661 '<|fim_suffix|>'
print_info: FIM MID token = 151660 '<|fim_middle|>'
print_info: FIM PAD token = 151662 '<|fim_pad|>'
print_info: FIM REP token = 151663 '<|repo_name|>'
print_info: FIM SEP token = 151664 '<|file_sep|>'
print_info: EOG token = 151643 '<|endoftext|>'
print_info: EOG token = 151645 '<|im_end|>'
print_info: EOG token = 151662 '<|fim_pad|>'
print_info: EOG token = 151663 '<|repo_name|>'
print_info: EOG token = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = false)
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors: CPU model buffer size = 0.00 MiB
load_tensors: ROCm0 model buffer size = 80561.98 MiB
load_tensors: ROCm_Host model buffer size = 315.30 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 4
llama_context: n_ctx = 65536
llama_context: n_ctx_seq = 65536
llama_context: n_batch = 2048
llama_context: n_ubatch = 2048
llama_context: causal_attn = 1
llama_context: flash_attn = enabled
llama_context: kv_unified = true
llama_context: freq_base = 10000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: ROCm_Host output buffer size = 2.32 MiB
llama_kv_cache: ROCm0 KV buffer size = 1536.00 MiB
llama_kv_cache: size = 1536.00 MiB ( 65536 cells, 12 layers, 4/1 seqs), K (f16): 768.00 MiB, V (f16): 768.00 MiB
llama_memory_recurrent: ROCm0 RS buffer size = 301.50 MiB
llama_memory_recurrent: size = 301.50 MiB ( 4 cells, 48 layers, 4 seqs), R (f32): 13.50 MiB, S (f32): 288.00 MiB
ggml_new_object: not enough space in the context's memory pool (needed 10711552, available 10711184)
/home/mark/llama.cpp/compile/llama.cpp-master/ggml/src/ggml.c:1679: GGML_ASSERT(obj_new) failed
[New LWP 105107]
[New LWP 105104]
[New LWP 105103]
[New LWP 105102]
[New LWP 105101]
[New LWP 105100]
[New LWP 105099]
[New LWP 105098]
[New LWP 105097]
[New LWP 105096]
[New LWP 105095]
[New LWP 105094]
[New LWP 105093]
[New LWP 105092]
[New LWP 105091]
[New LWP 105090]
[New LWP 105089]
[New LWP 105088]
[New LWP 105087]
[New LWP 105086]
[New LWP 105085]
[New LWP 105084]
[New LWP 105083]
[New LWP 105082]
[New LWP 105081]
[New LWP 105080]
[New LWP 105079]
[New LWP 105078]
[New LWP 105077]
[New LWP 105076]
[New LWP 105075]
[New LWP 105074]
[New LWP 105073]
[New LWP 105072]
[New LWP 105069]

This GDB supports auto-downloading debuginfo from the following URLs:
https://debuginfod.ubuntu.com
Enable debuginfod for this session? (y or [n]) [answered N; input not from terminal]
Debuginfod has been disabled.
To make this setting permanent, add 'set debuginfod enabled off' to .gdbinit.
[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: 没有那个文件或目录
#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 0x000071d8e32a013c in __internal_syscall_cancel (a1=, a2=, a3=, a4=, a5=0, a6=0, nr=61) at ./nptl/cancellation.c:49
warning: 49 ./nptl/cancellation.c: 没有那个文件或目录
#2 __syscall_cancel (a1=, a2=, a3=, a4=, a5=a5@entry=0,a6=a6@entry=0, nr=61) at ./nptl/cancellation.c:75
75 in ./nptl/cancellation.c
#3 0x000071d8e331c98f in __GI___wait4 (pid=, stat_loc=, options=, usage=) at ../sysdeps/unix/sysv/linux/wait4.c:30
warning: 30 ../sysdeps/unix/sysv/linux/wait4.c: 没有那个文件或目录
#4 0x000071d8e3cca9d3 in ggml_print_backtrace () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libggml-base.so.0
#5 0x000071d8e3ccab86 in ggml_abort () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libggml-base.so.0
#6 0x000071d8e3ccbae1 in ggml_new_tensor_impl.constprop () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libggml-base.so.0
#7 0x000071d8e3cd1b60 in ggml_view_4d () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libggml-base.so.0
#8 0x000071d8e3be145e in llm_build_qwen3next::build_delta_net_chunking(ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*, int) () from /home/mark/llama.cpp/compile/llama.cp p-master/build/bin/libllama.so.0
#9 0x000071d8e3be4130 in llm_build_qwen3next::build_layer_attn_linear(llm_graph_input_rs*, ggml_tensor*, ggml_tensor*, ggml_tensor*, int) () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libllama.so.0
#10 0x000071d8e3be467e in llm_build_qwen3next::llm_build_qwen3next(llama_model const&, llm_graph_params const&) () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libllama.so.0
#11 0x000071d8e3b18459 in llama_model::build_graph(llm_graph_params const&) const () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libllama.so.0
#12 0x000071d8e3aa5478 in llama_context::graph_reserve(unsigned int, unsigned int, unsigned int, llama_memory_context_i const*, bool) () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libllama.so.0
#13 0x000071d8e3aa83e9 in llama_context::llama_context(llama_model const&, llama_context_params) () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libllama.so.0
#14 0x000071d8e3aa8ef4 in llama_init_from_model () from /home/mark/llama.cpp/compile/llama.cpp-master/build/bin/libllama.so.0
#15 0x00005ba47b54e43a in common_init_from_params(common_params&) ()
#16 0x00005ba47b3dadd2 in server_context::load_model(common_params const&) ()
#17 0x00005ba47b3b77e2 in main ()
[Inferior 1 (process 105067) detached]
已中止 (核心已转储)

Originally posted by @2432896620-ctrl in #16095 (comment)

[wereretot]

I have the same memory issue, both with CUDA and CPU only.

[danielhanchen]

@pwilkin Sorry re putting the imatrix issue o the Qwen3-Next PR - you're correct <= 512 for ubatch-size works for imatrix - I should have provided more context sorry but as mentioned by others ubatch-size >= 513 gets the issue

[pwilkin]

There's apparently a problem with calculating the allocator size for batch sizes > 512.