Vulkan/AMDVLK: Device memory allocation fails when single compute buffer > ~2 GiB (same model works on RADV)

[kyuz0]

Problem description & steps to reproduce

There is an open issue on llama.ccp (ggml-org/llama.cpp#15054) for a bug on systems using AMDVLK. Some GGUF models that load and run under RADV fail during model/context initialization in llama.cpp’s Vulkan backend with VK_ERROR_OUT_OF_DEVICE_MEMORY:

ggml_vulkan: Device memory allocation of size 2819260416 failed.
ggml_vulkan: Requested buffer size exceeds device memory allocation limit: ErrorOutOfDeviceMemory
alloc_tensor_range: failed to allocate Vulkan0 buffer of size 2819260416

The failures correlate with a single Vulkan allocation (the ggml_vulkan compute buffer) exceeding the driver’s per-allocation cap VkPhysicalDeviceLimits::maxMemoryAllocationSize. On the affected machine this is 0x80000000 (≈2 GiB). The backend occasionally needs to request ~2.0–2.6 GiB for this buffer, which succeeds on RADV but is rejected by AMDVLK, indicating the issue is the per-allocation limit, not total VRAM.

The following response from the original llama.cpp issue made me think it was a good idea to raise this here:

Yes, the driver sets a maximum allocation size and a maximum buffer size limit. This is 2GB on amdvlk and the proprietary AMD drivers, and 4GB on RADV. We have no control over this, I don't know why AMD keeps the limit below the theoretical maximum that all the other major Vulkan drivers use.

Example (Gemma 3 27B BF16)

For example, this happens with gemma-3-27b-it-BF16/gemma-3-27b-it-BF16 (https://huggingface.co/unsloth/gemma-3-27b-it-GGUF):

$ llama-cli -ngl 99 -fa -m models/gemma-3-27b-it-BF16/gemma-3-27b-it-BF16-00001-of-00002.gguf 
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
build: 6060 (9c35706b) with cc (GCC) 15.1.1 20250719 (Red Hat 15.1.1-5) for x86_64-redhat-linux
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device Vulkan0 (Radeon 8060S Graphics) - 85720 MiB free
llama_model_loader: additional 1 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 39 key-value pairs and 808 tensors from models/gemma-3-27b-it-BF16/gemma-3-27b-it-BF16-00001-of-00002.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              = gemma3
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Gemma-3-27B-It
llama_model_loader: - kv   3:                           general.finetune str              = it
llama_model_loader: - kv   4:                           general.basename str              = Gemma-3-27B-It
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 27B
llama_model_loader: - kv   7:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   8:                      gemma3.context_length u32              = 131072
llama_model_loader: - kv   9:                    gemma3.embedding_length u32              = 5376
llama_model_loader: - kv  10:                         gemma3.block_count u32              = 62
llama_model_loader: - kv  11:                 gemma3.feed_forward_length u32              = 21504
llama_model_loader: - kv  12:                gemma3.attention.head_count u32              = 32
llama_model_loader: - kv  13:    gemma3.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:                gemma3.attention.key_length u32              = 128
llama_model_loader: - kv  15:              gemma3.attention.value_length u32              = 128
llama_model_loader: - kv  16:                          general.file_type u32              = 32
llama_model_loader: - kv  17:                      gemma3.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  18:            gemma3.attention.sliding_window u32              = 1024
llama_model_loader: - kv  19:             gemma3.attention.head_count_kv u32              = 16
llama_model_loader: - kv  20:                   gemma3.rope.scaling.type str              = linear
llama_model_loader: - kv  21:                 gemma3.rope.scaling.factor f32              = 8.000000
llama_model_loader: - kv  22:               general.quantization_version u32              = 2
llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  24:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  25:                      tokenizer.ggml.tokens arr[str,262208]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  26:                      tokenizer.ggml.scores arr[f32,262208]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  27:                  tokenizer.ggml.token_type arr[i32,262208]  = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  28:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  29:                tokenizer.ggml.eos_token_id u32              = 106
llama_model_loader: - kv  30:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  31:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  32:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  33:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {{ bos_token }}\n{%- if messages[0]['r...
llama_model_loader: - kv  35:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  36:                                   split.no u16              = 0
llama_model_loader: - kv  37:                                split.count u16              = 2
llama_model_loader: - kv  38:                        split.tensors.count i32              = 808
llama_model_loader: - type  f32:  373 tensors
llama_model_loader: - type bf16:  435 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = BF16
print_info: file size   = 50.31 GiB (16.00 BPW) 
load: special tokens cache size = 6415
load: token to piece cache size = 1.9446 MB
print_info: arch             = gemma3
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 5376
print_info: n_layer          = 62
print_info: n_head           = 32
print_info: n_head_kv        = 16
print_info: n_rot            = 128
print_info: n_swa            = 1024
print_info: is_swa_any       = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 2
print_info: n_embd_k_gqa     = 2048
print_info: n_embd_v_gqa     = 2048
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     = 7.7e-02
print_info: n_ff             = 21504
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  = 1000000.0
print_info: freq_scale_train = 0.125
print_info: n_ctx_orig_yarn  = 131072
print_info: rope_finetuned   = unknown
print_info: model type       = 27B
print_info: model params     = 27.01 B
print_info: general.name     = Gemma-3-27B-It
print_info: vocab type       = SPM
print_info: n_vocab          = 262208
print_info: n_merges         = 0
print_info: BOS token        = 2 '<bos>'
print_info: EOS token        = 106 '<end_of_turn>'
print_info: EOT token        = 106 '<end_of_turn>'
print_info: UNK token        = 3 '<unk>'
print_info: PAD token        = 0 '<pad>'
print_info: LF token         = 248 '<0x0A>'
print_info: EOG token        = 106 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
ggml_vulkan: Device memory allocation of size 2819260416 failed.
ggml_vulkan: Requested buffer size exceeds device memory allocation limit: ErrorOutOfDeviceMemory
alloc_tensor_range: failed to allocate Vulkan0 buffer of size 2819260416
llama_model_load: error loading model: unable to allocate Vulkan0 buffer
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model 'models/gemma-3-27b-it-BF16/gemma-3-27b-it-BF16-00001-of-00002.gguf'
main: error: unable to load model

However, with RADV the model loads just fine:

toolbox enter llama-vulkan-radv 
⬢ [kyuz0@toolbx ~]$ llama-cli -ngl 99 -fa -m models/gemma-3-27b-it-BF16/gemma-3-27b-it-BF16-00001-of-00002.gguf 
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
build: 6040 (66625a59) with cc (GCC) 15.1.1 20250719 (Red Hat 15.1.1-5) for x86_64-redhat-linux
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device Vulkan0 (Radeon 8060S Graphics (RADV GFX1151)) - 87722 MiB free
llama_model_loader: additional 1 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 39 key-value pairs and 808 tensors from models/gemma-3-27b-it-BF16/gemma-3-27b-it-BF16-00001-of-00002.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              = gemma3
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Gemma-3-27B-It
llama_model_loader: - kv   3:                           general.finetune str              = it
llama_model_loader: - kv   4:                           general.basename str              = Gemma-3-27B-It
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 27B
llama_model_loader: - kv   7:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   8:                      gemma3.context_length u32              = 131072
llama_model_loader: - kv   9:                    gemma3.embedding_length u32              = 5376
llama_model_loader: - kv  10:                         gemma3.block_count u32              = 62
llama_model_loader: - kv  11:                 gemma3.feed_forward_length u32              = 21504
llama_model_loader: - kv  12:                gemma3.attention.head_count u32              = 32
llama_model_loader: - kv  13:    gemma3.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:                gemma3.attention.key_length u32              = 128
llama_model_loader: - kv  15:              gemma3.attention.value_length u32              = 128
llama_model_loader: - kv  16:                          general.file_type u32              = 32
llama_model_loader: - kv  17:                      gemma3.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  18:            gemma3.attention.sliding_window u32              = 1024
llama_model_loader: - kv  19:             gemma3.attention.head_count_kv u32              = 16
llama_model_loader: - kv  20:                   gemma3.rope.scaling.type str              = linear
llama_model_loader: - kv  21:                 gemma3.rope.scaling.factor f32              = 8.000000
llama_model_loader: - kv  22:               general.quantization_version u32              = 2
llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  24:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  25:                      tokenizer.ggml.tokens arr[str,262208]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  26:                      tokenizer.ggml.scores arr[f32,262208]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  27:                  tokenizer.ggml.token_type arr[i32,262208]  = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  28:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  29:                tokenizer.ggml.eos_token_id u32              = 106
llama_model_loader: - kv  30:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  31:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  32:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  33:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {{ bos_token }}\n{%- if messages[0]['r...
llama_model_loader: - kv  35:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  36:                                   split.no u16              = 0
llama_model_loader: - kv  37:                                split.count u16              = 2
llama_model_loader: - kv  38:                        split.tensors.count i32              = 808
llama_model_loader: - type  f32:  373 tensors
llama_model_loader: - type bf16:  435 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = BF16
print_info: file size   = 50.31 GiB (16.00 BPW) 
load: special tokens cache size = 6415
load: token to piece cache size = 1.9446 MB
print_info: arch             = gemma3
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 5376
print_info: n_layer          = 62
print_info: n_head           = 32
print_info: n_head_kv        = 16
print_info: n_rot            = 128
print_info: n_swa            = 1024
print_info: is_swa_any       = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 2
print_info: n_embd_k_gqa     = 2048
print_info: n_embd_v_gqa     = 2048
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     = 7.7e-02
print_info: n_ff             = 21504
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  = 1000000.0
print_info: freq_scale_train = 0.125
print_info: n_ctx_orig_yarn  = 131072
print_info: rope_finetuned   = unknown
print_info: model type       = 27B
print_info: model params     = 27.01 B
print_info: general.name     = Gemma-3-27B-It
print_info: vocab type       = SPM
print_info: n_vocab          = 262208
print_info: n_merges         = 0
print_info: BOS token        = 2 '<bos>'
print_info: EOS token        = 106 '<end_of_turn>'
print_info: EOT token        = 106 '<end_of_turn>'
print_info: UNK token        = 3 '<unk>'
print_info: PAD token        = 0 '<pad>'
print_info: LF token         = 248 '<0x0A>'
print_info: EOG token        = 106 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 62 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 63/63 layers to GPU
load_tensors:      Vulkan0 model buffer size = 51518.82 MiB
load_tensors:   CPU_Mapped model buffer size =  2688.66 MiB
.............................................................................................
llama_context: constructing llama_context
llama_context: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 0.125
llama_context: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     1.00 MiB
llama_kv_cache_unified_iswa: creating non-SWA KV cache, size = 4096 cells
llama_kv_cache_unified:    Vulkan0 KV buffer size =   320.00 MiB
llama_kv_cache_unified: size =  320.00 MiB (  4096 cells,  10 layers,  1/ 1 seqs), K (f16):  160.00 MiB, V (f16):  160.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_kv_cache_unified_iswa: creating     SWA KV cache, size = 1536 cells
llama_kv_cache_unified:    Vulkan0 KV buffer size =   624.00 MiB
llama_kv_cache_unified: size =  624.00 MiB (  1536 cells,  52 layers,  1/ 1 seqs), K (f16):  312.00 MiB, V (f16):  312.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_context:    Vulkan0 compute buffer size =   522.62 MiB
llama_context: Vulkan_Host compute buffer size =    21.51 MiB
llama_context: graph nodes  = 2613
llama_context: graph splits = 2
common_init_from_params: KV cache shifting is not supported for this context, disabling KV cache shifting
common_init_from_params: added <end_of_turn> logit bias = -inf
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)
main: llama threadpool init, n_threads = 16
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
<start_of_turn>user
You are a helpful assistant

Hello<end_of_turn>
<start_of_turn>model
Hi there<end_of_turn>
<start_of_turn>user
How are you?<end_of_turn>
<start_of_turn>model


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

main: interactive mode on.
sampler seed: 3532905611
sampler params: 
	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
	dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 4096
	top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist 
generate: n_ctx = 4096, n_batch = 2048, n_predict = -1, n_keep = 1

== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to the AI.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.
 - Not using system message. To change it, set a different value via -sys PROMPT

Vulkaninfo output

$vulkaninfo | grep -i maxMemoryAllocationSize

'DISPLAY' environment variable not set... skipping surface info
	maxMemoryAllocationSize           = 0xfffffffc
	maxMemoryAllocationSize           = 0x80000000

$vulkaninfo | grep -i maxMemoryAllocationSize

'DISPLAY' environment variable not set... skipping surface info
	maxMemoryAllocationSize           = 0x80000000

Name and Version

llama-cli --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon 8060S Graphics (AMD open-source driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
version: 6060 (9c35706b)
built with cc (GCC) 15.1.1 20250719 (Red Hat 15.1.1-5) for x86_64-redhat-linux

Operating systems

Linux

[lhl]

Just an FYI, I was taking a look at this just now and it seems the main issue is that ggml-org/llama.cpp#15120 (comment)

[mike-fischer-ml]

For the use case of local LLM's , if its not possible allocate enough memory to the vulkan amd device, the AMD gpu/apu's really can't run any decent local model.

If you read the llama cpp github issues, theres a decent of amount problems related to memory allocations.

Might want to spend some time on this.

[visorcraft]

Adding a related gfx1151 report focused on capability/perf mismatch (not only allocation-limit failures):

• New issue: gfx1151 (Radeon 8060S / Strix Halo): Vulkan capability/perf mismatch vs RADV in llama.cpp workloads #420
  gfx1151 (Radeon 8060S / Strix Halo): Vulkan capability/perf mismatch vs RADV in llama.cpp workloads #420

On Ryzen AI Max+ 395 / Radeon 8060S (gfx1151), we see large llama.cpp Vulkan prompt-processing regression when using the AMD open-source/LLPC path vs RADV on identical hardware/model/flags.

Key signal:

• AMD open-source path reports shared memory: 32768
• RADV path reports shared memory: 65536

This correlates with pp512 moving from ~378 (AMD path) to ~507–522 (RADV path) for Qwen3-Coder-Next 80B-A3B Q4_K_M.

Keeping #413 for allocation-limit behavior, cross-linking #420 because both may be part of the same gfx1151 AMDVLK Vulkan limitation family.