Ollama: Crash on Intel Iris Xe due to non-configurable pipeline parallelism (n_copies=4)

[jaredpilcher]

🐛 Describe the bug

When running models like devstral:latest using the ipex-llm build of Ollama, the runner process crashes with an exit status 1. The server log consistently shows that pipeline parallelism enabled (n_copies=4) is active by default. This aggressive parallelism setting appears to overwhelm the resources of integrated GPUs like the Intel Iris Xe, causing the Intel SYCL driver to crash with a UR errorException.

This n_copies=4 setting cannot be configured. The num_parallel parameter is not recognized in the Modelfile for this version of Ollama, making it impossible to reduce the parallelism and prevent the crash.

To Reproduce

1. Install ipex-llm[cpp] on a Windows machine with an Intel Iris Xe GPU.
2. Start the Ollama server: ollama serve
3. Run a large model: ollama run devstral:latest
4. The process will begin to load, allocate buffers, and then crash.

Expected behavior

The model should load and run successfully without crashing the driver. Ideally, the default n_copies for integrated GPUs would be lower (e.g., 1 or 2), or there should be a supported parameter in the Modelfile (like num_parallel) to configure it manually.

Environment

• OS: Windows
• Hardware: Lenovo ThinkPad T14s Gen 4
• GPU: Intel(R) Iris(R) Xe Graphics
• Library: ipex-llm[cpp]

Attempts to Troubleshoot

• Setting num_gpu to a low value (e.g., 10) in a Modelfile does not solve the issue, as the n_copies=4 setting still triggers the crash.
• Adding PARAMETER num_parallel 1 to a Modelfile results in an Error: unknown parameter 'num_parallel', confirming this version does not support the standard Ollama parameter for controlling parallelism.

Ollama Server Log

llama_model_loader: - type  f32:   81 tensors
llama_model_loader: - type q4_K:  241 tensors
llama_model_loader: - type q6_K:   41 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 13.34 GiB (4.86 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 1000
load: token to piece cache size = 0.8498 MB
print_info: arch             = llama
print_info: vocab_only       = 1
print_info: model type       = ?B
print_info: model params     = 23.57 B
print_info: general.name     = Devstral Small 2505
print_info: vocab type       = BPE
print_info: n_vocab          = 131072
print_info: n_merges         = 269443
print_info: BOS token        = 1 '<s>'
print_info: EOS token        = 2 '</s>'
print_info: UNK token        = 0 '<unk>'
print_info: LF token         = 1010 'Ċ'
print_info: EOG token        = 2 '</s>'
print_info: max token length = 150
llama_model_load: vocab only - skipping tensors
time=2025-06-28T23:37:14.961-06:00 level=INFO source=server.go:430 msg="starting llama server" cmd="C:\\Users\\pilchj\\AppData\\Local\\miniconda3\\envs\\llm-cpp\\Lib\\site-packages\\bigdl\\cpp\\libs\\ollama\\ollama-lib.exe runner --model C:\\Users\\pilchj\\.ollama\\models\\blobs\\sha256-b3a2c9a8fef9be8d2ef951aecca36a36b9ea0b70abe9359eab4315bf4cd9be01 --ctx-size 8192 --batch-size 512 --n-gpu-layers 999 --threads 6 --no-mmap --parallel 4 --port 50155"
time=2025-06-28T23:37:14.984-06:00 level=INFO source=sched.go:450 msg="loaded runners" count=1
time=2025-06-28T23:37:14.994-06:00 level=INFO source=server.go:605 msg="waiting for llama runner to start responding"
time=2025-06-28T23:37:14.996-06:00 level=INFO source=server.go:639 msg="waiting for server to become available" status="llm server error"
using override patterns: []
time=2025-06-28T23:37:15.256-06:00 level=INFO source=runner.go:883 msg="starting go runner"
time=2025-06-28T23:37:15.266-06:00 level=INFO source=ggml.go:109 msg=system CPU.0.LLAMAFILE=1 compiler=cgo(clang)
ModelParams: {NumGpuLayers:999 MainGpu:0 UseMmap:false UseMlock:false TensorSplit:[] Progress:0x7ff7e1aee200 VocabOnly:false OverrideTensors:[]}
time=2025-06-28T23:37:15.269-06:00 level=INFO source=runner.go:944 msg="Server listening on 127.0.0.1:50155"
time=2025-06-28T23:37:15.502-06:00 level=INFO source=server.go:639 msg="waiting for server to become available" status="llm server loading model"
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
llama_model_load_from_file_impl: using device SYCL0 (Intel(R) Iris(R) Xe Graphics) - 14789 MiB free
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
llama_model_load_from_file_impl: using device SYCL0 (Intel(R) Iris(R) Xe Graphics) - 14789 MiB free
llama_model_loader: loaded meta data with 41 key-value pairs and 363 tensors from C:\Users\pilchj\.ollama\models\blobs\sha256-b3a2c9a8fef9be8d2ef951aecca36a36b9ea0b70abe9359eab4315bf4cd9be01 (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              = llama
llama_model_loader: - kv   1:                  general.base_model.0.name str              = Devstrall Small 2505
llama_model_loader: - kv   2:          general.base_model.0.organization str              = Mistralai
llama_model_loader: - kv   3:              general.base_model.0.repo_url str              = https://huggingface.co/mistralai/Devs...
llama_model_loader: - kv   4:               general.base_model.0.version str              = 2505
llama_model_loader: - kv   5:                   general.base_model.count u32              = 1
llama_model_loader: - kv   6:                           general.basename str              = Devstral
llama_model_loader: - kv   7:                          general.file_type u32              = 15
llama_model_loader: - kv   8:                          general.languages arr[str,24]      = ["en", "fr", "de", "es", "pt", "it", ...
llama_model_loader: - kv   9:                            general.license str              = apache-2.0
llama_model_loader: - kv  10:                               general.name str              = Devstral Small 2505
llama_model_loader: - kv  11:                    general.parameter_count u64              = 23572403200
llama_model_loader: - kv  12:               general.quantization_version u32              = 2
llama_model_loader: - kv  13:                         general.size_label str              = Small
llama_model_loader: - kv  14:                               general.tags arr[str,1]       = ["text2text-generation"]
llama_model_loader: - kv  15:                               general.type str              = model
llama_model_loader: - kv  16:                            general.version str              = 2505
llama_model_loader: - kv  17:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  18:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  19:                 llama.attention.key_length u32              = 128
llama_model_loader: - kv  20:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  21:               llama.attention.value_length u32              = 128
llama_model_loader: - kv  22:                          llama.block_count u32              = 40
llama_model_loader: - kv  23:                       llama.context_length u32              = 131072
llama_model_loader: - kv  24:                     llama.embedding_length u32              = 5120
llama_model_loader: - kv  25:                  llama.feed_forward_length u32              = 32768
llama_model_loader: - kv  26:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  27:                       llama.rope.freq_base f32              = 1000000000.000000
llama_model_loader: - kv  28:                           llama.vocab_size u32              = 131072
llama_model_loader: - kv  29:                    tokenizer.chat_template str              = {%- set today = strftime_now("%Y-%m-%...
llama_model_loader: - kv  30:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  31:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  32:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  33:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  34:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  35:                      tokenizer.ggml.merges arr[str,269443]  = ["Ġ Ġ", "Ġ t", "e r", "i n", "Ġ �...
llama_model_loader: - kv  36:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  37:                         tokenizer.ggml.pre str              = tekken
llama_model_loader: - kv  38:                  tokenizer.ggml.token_type arr[i32,131072]  = [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  39:                      tokenizer.ggml.tokens arr[str,131072]  = ["<unk>", "<s>", "</s>", "[INST]", "[...
llama_model_loader: - kv  40:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - type  f32:   81 tensors
llama_model_loader: - type q4_K:  241 tensors
llama_model_loader: - type q6_K:   41 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 13.34 GiB (4.86 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 1000
load: token to piece cache size = 0.8498 MB
print_info: arch             = llama
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 40
print_info: n_head           = 32
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
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     = 1024
print_info: n_embd_v_gqa     = 1024
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: n_ff             = 32768
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        = 0
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 131072
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       = 13B
print_info: model params     = 23.57 B
print_info: general.name     = Devstral Small 2505
print_info: vocab type       = BPE
print_info: n_vocab          = 131072
print_info: n_merges         = 269443
print_info: BOS token        = 1 '<s>'
print_info: EOS token        = 2 '</s>'
print_info: UNK token        = 0 '<unk>'
print_info: LF token         = 1010 'Ċ'
print_info: EOG token        = 2 '</s>'
print_info: max token length = 150
load_tensors: loading model tensors, this can take a while... (mmap = false)
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
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:        SYCL0 model buffer size = 13302.36 MiB
load_tensors:          CPU model buffer size =   360.00 MiB
llama_init_from_model: n_seq_max     = 4
llama_init_from_model: n_ctx         = 8192
llama_init_from_model: n_ctx_per_seq = 2048
llama_init_from_model: n_batch       = 2048
llama_init_from_model: n_ubatch      = 512
llama_init_from_model: flash_attn    = 0
llama_init_from_model: freq_base     = 1000000000.0
llama_init_from_model: freq_scale    = 1
llama_init_from_model: n_ctx_per_seq (2048) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
Running with Environment Variables:
  GGML_SYCL_DEBUG: 0
  GGML_SYCL_DISABLE_OPT: 1
Build with Macros:
  GGML_SYCL_FORCE_MMQ: no
  GGML_SYCL_F16: no
Found 1 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]|                 Intel Iris Xe Graphics|   12.3|     96|     512|   32| 15507M|            1.6.32960|
SYCL Optimization Feature:
|ID|        Device Type|Reorder|
|--|-------------------|-------|
| 0| [level_zero:gpu:0]|      N|
llama_kv_cache_init: kv_size = 8192, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 40, can_shift = 1
llama_kv_cache_init:      SYCL0 KV buffer size =  1280.00 MiB
llama_init_from_model: KV self size  = 1280.00 MiB, K (f16):  640.00 MiB, V (f16):  640.00 MiB
llama_init_from_model:  SYCL_Host  output buffer size =     2.08 MiB
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
llama_init_from_model: pipeline parallelism enabled (n_copies=4)
llama_init_from_model:      SYCL0 compute buffer size =   410.02 MiB
llama_init_from_model:  SYCL_Host compute buffer size =    74.02 MiB
llama_init_from_model: graph nodes  = 1126 (with bs=512), 1006 (with bs=1)
llama_init_from_model: graph splits = 3
time=2025-06-28T23:37:57.940-06:00 level=WARN source=runner.go:802 msg="%s: warming up the model with an empty run - please wait ... " !BADKEY=loadModel
UR errorException caught at file:D:\actions-runner\release-cpp-oneapi_2024_2\_work\llm.cpp\llm.cpp\ollama-llama-cpp\ggml\src\ggml-sycl\common.cpp, line:99
time=2025-06-28T23:37:59.880-06:00 level=INFO source=server.go:639 msg="waiting for server to become available" status="llm server not responding"
time=2025-06-28T23:38:01.936-06:00 level=INFO source=server.go:639 msg="waiting for server to become available" status="llm server error"
time=2025-06-28T23:38:02.187-06:00 level=ERROR source=sched.go:456 msg="error loading llama server" error="llama runner process has terminated: exit status 1"
[GIN] 2025/06/28 - 23:38:02 | 500 |   48.2099198s |       127.0.0.1 | POST     "/api/generate"

[rnwang04]

Hi @jaredpilcher , if you want to reduce num_parallel to save kv cache memory, you could set OLLAMA_NUM_PARALLEL=1 in your cmd.