github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-11-12 03:46:14 -06:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #12237] dolphin-llama3 256K context window support is broken #33901

New Issue
Closed
opened 2026-04-22 17:05:13 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-22 17:05:13 -05:00
Owner
Copy Link
Copy Source

Originally created by @ruro on GitHub (Sep 10, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12237

What is the issue?

I am not sure if this is the right place to report this issue, but the official model page for dolphin-llama3 claims that dolphin-llama3 supports extending the context window to 256K tokens.

However, this doesn't actually seem to work. For example, consider

import ollama
import transformers

# estimate the tokenized message length
tokenizer = transformers.AutoTokenizer.from_pretrained("dphn/dolphin-2.9-llama3-70b")
def get_approx_num_tokens(text): return len(tokenizer(text)["input_ids"])

# creates a prompt suitable for testing the context window size
def make_context_length_test(num_repeats):
    text = [
        "I am testing the context length of this model. Please, remember the following number: 57",
        "Ignore all following text until I say 'CONTINUE'.",
        " ".join(num_repeats * ["IGNORE THIS"]),
        "CONTINUE",
        "Earlier I asked you to remember a number. What was that number?",
    ]
    return "\n\n".join(text)

for num_ctx, num_repeats in [
    (None, 2000),
    (None, 2500),

    (8192, 4000),
    (8192, 4500),

    (256000, 4500),
]:
    test = make_context_length_test(num_repeats)
    num_tokens = get_approx_num_tokens(test)
    print(f"Testing {num_ctx=}, {num_tokens=}...")
    print(
        "Model response:",
        ollama.chat(
            model="dolphin-llama3:8b",
            messages=[{"role": "user", "content": test}],
            options=ollama.Options(num_ctx=num_ctx, seed=42),
        ).message.content.strip()
    )
    print()

This code runs dolphin-llama3:8b with a few different num_ctx values. The user asks the model to remember the number, then repeats "IGNORE THIS" a bunch and then asks the model what number it was given.

Here is an example of the outputs:

Testing num_ctx=None, num_tokens=4050...
Model response: 57

Testing num_ctx=None, num_tokens=5050...
Model response: Apologies, I am unable to recall the specific number you are referring to without any context or further details provided in our previous conversation. Could you please provide more information or clarify the situation for me?

Testing num_ctx=8192, num_tokens=8050...
Model response: The number you wanted me to remember was 57.

Testing num_ctx=8192, num_tokens=9050...
Model response: To help you remember the number, let me know which one it is and I'll try my best to describe it for you.

Testing num_ctx=256000, num_tokens=9050...
Model response: To help you remember the number, let me know which one it is. If it's an important series of events or steps, we can try to organize them into smaller parts, making each part shorter and less complex than the whole. Breaking it down will make it easier for you to recall later.

Or in other words. With the default num_ctx the model seems to give the correct response for ~4000 tokens, but fails for ~5000 tokens. Raising num_ctx to 8192 makes the model remember the number for up to ~8000 tokens.

However, trying to further raise the num_ctx to 256000 (as recommended by the model page) doesn't actually seem to increase the context window any further. The model still fails for inputs larger than ~8000 tokens.

Looking at the logs of ollama serve, the following lines seem to be the most important:

time=2025-09-10T05:38:12.651+03:00 level=WARN source=server.go:145 msg="requested context size too large for model" num_ctx=256000 num_parallel=1 n_ctx_train=8192
llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
print_info: n_ctx_train      = 8192
llama_context: n_ctx         = 8192
time=2025-09-10T05:38:13.658+03:00 level=WARN source=runner.go:128 msg="truncating input prompt" limit=8192 prompt=9078 keep=4 new=8192

So it looks like something somewhere inside ollama silently forces the context length to be limited to 8192.

Relevant log output

time=2025-09-10T05:38:05.155+03:00 level=INFO source=routes.go:1297 msg="server config" env="map[CUDA_VISIBLE_DEVICES: GPU_DEVICE_ORDINAL: HIP_VISIBLE_DEVICES: HSA_OVERRIDE_GFX_VERSION: HTTPS_PROXY: HTTP_PROXY: NO_PROXY: OLLAMA_CONTEXT_LENGTH:4096 OLLAMA_DEBUG:INFO OLLAMA_FLASH_ATTENTION:true OLLAMA_GPU_OVERHEAD:0 OLLAMA_HOST:http://127.0.0.1:11434 OLLAMA_INTEL_GPU:false OLLAMA_KEEP_ALIVE:24h0m0s OLLAMA_KV_CACHE_TYPE: OLLAMA_LLM_LIBRARY: OLLAMA_LOAD_TIMEOUT:5m0s OLLAMA_MAX_LOADED_MODELS:0 OLLAMA_MAX_QUEUE:512 OLLAMA_MODELS:/home/ollama/.ollama/models OLLAMA_MULTIUSER_CACHE:false OLLAMA_NEW_ENGINE:false OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:1 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://* vscode-webview://* vscode-file://*] OLLAMA_SCHED_SPREAD:false ROCR_VISIBLE_DEVICES: http_proxy: https_proxy: no_proxy:]"
time=2025-09-10T05:38:05.161+03:00 level=INFO source=images.go:477 msg="total blobs: 60"
time=2025-09-10T05:38:05.163+03:00 level=INFO source=images.go:484 msg="total unused blobs removed: 0"
time=2025-09-10T05:38:05.164+03:00 level=INFO source=routes.go:1350 msg="Listening on 127.0.0.1:11434 (version 0.11.3)"
time=2025-09-10T05:38:05.165+03:00 level=INFO source=gpu.go:217 msg="looking for compatible GPUs"
time=2025-09-10T05:38:05.400+03:00 level=INFO source=types.go:130 msg="inference compute" id=GPU-25c74f4d-1222-c2f3-5f38-9e1105148d68 library=cuda variant=v12 compute=12.0 driver=12.8 name="NVIDIA GeForce RTX 5070 Ti" total="15.5 GiB" available="15.2 GiB"
time=2025-09-10T05:38:12.651+03:00 level=INFO source=server.go:135 msg="system memory" total="125.4 GiB" free="74.8 GiB" free_swap="0 B"
time=2025-09-10T05:38:12.651+03:00 level=WARN source=server.go:145 msg="requested context size too large for model" num_ctx=256000 num_parallel=1 n_ctx_train=8192
time=2025-09-10T05:38:12.751+03:00 level=INFO source=server.go:175 msg=offload library=cuda layers.requested=-1 layers.model=33 layers.offload=33 layers.split="" memory.available="[15.2 GiB]" memory.gpu_overhead="0 B" memory.required.full="6.2 GiB" memory.required.partial="6.2 GiB" memory.required.kv="1.0 GiB" memory.required.allocations="[6.2 GiB]" memory.weights.total="4.1 GiB" memory.weights.repeating="3.7 GiB" memory.weights.nonrepeating="411.0 MiB" memory.graph.full="560.0 MiB" memory.graph.partial="677.5 MiB"
time=2025-09-10T05:38:12.752+03:00 level=INFO source=server.go:218 msg="enabling flash attention"
time=2025-09-10T05:38:12.752+03:00 level=WARN source=server.go:226 msg="kv cache type not supported by model" type=""
llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from /home/ollama/.ollama/models/blobs/sha256-ea025c107c1c3e5a87380b25e205db3b19a4d2ed56f1954f531573a963621580 (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.name str              = dolphin-2.9-llama3-8b
llama_model_loader: - kv   2:                          llama.block_count u32              = 32
llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 2
llama_model_loader: - kv  11:                           llama.vocab_size u32              = 128258
llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                      tokenizer.ggml.tokens arr[str,128258]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,128258]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  16:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 128256
llama_model_loader: - kv  19:            tokenizer.ggml.padding_token_id u32              = 128001
llama_model_loader: - kv  20:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q4_0:  225 tensors
llama_model_loader: - type q6_K:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_0
print_info: file size   = 4.33 GiB (4.64 BPW) 
load: missing or unrecognized pre-tokenizer type, using: 'default'
load: special tokens cache size = 258
load: token to piece cache size = 0.8000 MB
print_info: arch             = llama
print_info: vocab_only       = 1
print_info: model type       = ?B
print_info: model params     = 8.03 B
print_info: general.name     = dolphin-2.9-llama3-8b
print_info: vocab type       = BPE
print_info: n_vocab          = 128258
print_info: n_merges         = 280147
print_info: BOS token        = 128000 '<|begin_of_text|>'
print_info: EOS token        = 128256 '<|im_end|>'
print_info: EOT token        = 128256 '<|im_end|>'
print_info: PAD token        = 128001 '<|end_of_text|>'
print_info: LF token         = 198 'Ċ'
print_info: EOG token        = 128009 '<|eot_id|>'
print_info: EOG token        = 128256 '<|im_end|>'
print_info: max token length = 256
llama_model_load: vocab only - skipping tensors
time=2025-09-10T05:38:12.869+03:00 level=INFO source=server.go:438 msg="starting llama server" cmd="/opt/ollama/bin/ollama runner --model /home/ollama/.ollama/models/blobs/sha256-ea025c107c1c3e5a87380b25e205db3b19a4d2ed56f1954f531573a963621580 --ctx-size 8192 --batch-size 512 --n-gpu-layers 33 --threads 8 --flash-attn --parallel 1 --port 35583"
time=2025-09-10T05:38:12.869+03:00 level=INFO source=sched.go:481 msg="loaded runners" count=1
time=2025-09-10T05:38:12.869+03:00 level=INFO source=server.go:598 msg="waiting for llama runner to start responding"
time=2025-09-10T05:38:12.869+03:00 level=INFO source=server.go:632 msg="waiting for server to become available" status="llm server not responding"
time=2025-09-10T05:38:12.875+03:00 level=INFO source=runner.go:815 msg="starting go runner"
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 5070 Ti, compute capability 12.0, VMM: yes
load_backend: loaded CUDA backend from /opt/ollama/lib/ollama/libggml-cuda.so
load_backend: loaded CPU backend from /opt/ollama/lib/ollama/libggml-cpu-alderlake.so
time=2025-09-10T05:38:12.913+03:00 level=INFO source=ggml.go:104 msg=system CPU.0.SSE3=1 CPU.0.SSSE3=1 CPU.0.AVX=1 CPU.0.AVX_VNNI=1 CPU.0.AVX2=1 CPU.0.F16C=1 CPU.0.FMA=1 CPU.0.BMI2=1 CPU.0.LLAMAFILE=1 CPU.1.LLAMAFILE=1 CUDA.0.ARCHS=750,800,860,890,900,1000,1200 CUDA.0.USE_GRAPHS=1 CUDA.0.PEER_MAX_BATCH_SIZE=128 compiler=cgo(gcc)
time=2025-09-10T05:38:12.913+03:00 level=INFO source=runner.go:874 msg="Server listening on 127.0.0.1:35583"
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 5070 Ti) - 15602 MiB free
llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from /home/ollama/.ollama/models/blobs/sha256-ea025c107c1c3e5a87380b25e205db3b19a4d2ed56f1954f531573a963621580 (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.name str              = dolphin-2.9-llama3-8b
llama_model_loader: - kv   2:                          llama.block_count u32              = 32
llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 2
llama_model_loader: - kv  11:                           llama.vocab_size u32              = 128258
llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                      tokenizer.ggml.tokens arr[str,128258]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,128258]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  16:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 128256
llama_model_loader: - kv  19:            tokenizer.ggml.padding_token_id u32              = 128001
llama_model_loader: - kv  20:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q4_0:  225 tensors
llama_model_loader: - type q6_K:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_0
print_info: file size   = 4.33 GiB (4.64 BPW) 
load: missing or unrecognized pre-tokenizer type, using: 'default'
load: special tokens cache size = 258
time=2025-09-10T05:38:13.120+03:00 level=INFO source=server.go:632 msg="waiting for server to become available" status="llm server loading model"
load: token to piece cache size = 0.8000 MB
print_info: arch             = llama
print_info: vocab_only       = 0
print_info: n_ctx_train      = 8192
print_info: n_embd           = 4096
print_info: n_layer          = 32
print_info: n_head           = 32
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
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: f_attn_scale     = 0.0e+00
print_info: n_ff             = 14336
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  = 500000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 8192
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       = 8B
print_info: model params     = 8.03 B
print_info: general.name     = dolphin-2.9-llama3-8b
print_info: vocab type       = BPE
print_info: n_vocab          = 128258
print_info: n_merges         = 280147
print_info: BOS token        = 128000 '<|begin_of_text|>'
print_info: EOS token        = 128256 '<|im_end|>'
print_info: EOT token        = 128256 '<|im_end|>'
print_info: PAD token        = 128001 '<|end_of_text|>'
print_info: LF token         = 198 'Ċ'
print_info: EOG token        = 128009 '<|eot_id|>'
print_info: EOG token        = 128256 '<|im_end|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 32 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 33/33 layers to GPU
load_tensors:        CUDA0 model buffer size =  4156.00 MiB
load_tensors:   CPU_Mapped model buffer size =   281.82 MiB
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 8192
llama_context: n_ctx_per_seq = 8192
llama_context: n_batch       = 512
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 500000.0
llama_context: freq_scale    = 1
llama_context:  CUDA_Host  output buffer size =     0.50 MiB
llama_kv_cache_unified: kv_size = 8192, type_k = 'f16', type_v = 'f16', n_layer = 32, can_shift = 1, padding = 256
llama_kv_cache_unified:      CUDA0 KV buffer size =  1024.00 MiB
llama_kv_cache_unified: KV self size  = 1024.00 MiB, K (f16):  512.00 MiB, V (f16):  512.00 MiB
llama_context:      CUDA0 compute buffer size =   258.50 MiB
llama_context:  CUDA_Host compute buffer size =    24.01 MiB
llama_context: graph nodes  = 967
llama_context: graph splits = 2
time=2025-09-10T05:38:13.622+03:00 level=INFO source=server.go:637 msg="llama runner started in 0.75 seconds"
time=2025-09-10T05:38:13.658+03:00 level=WARN source=runner.go:128 msg="truncating input prompt" limit=8192 prompt=9078 keep=4 new=8192

OS

Linux

GPU

Nvidia

CPU

Intel

Ollama version

0.11.3

Originally created by @ruro on GitHub (Sep 10, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12237 ### What is the issue? I am not sure if this is the right place to report this issue, but the official [model page for `dolphin-llama3`](https://ollama.com/library/dolphin-llama3) claims that `dolphin-llama3` supports extending the context window to 256K tokens. However, this doesn't actually seem to work. For example, consider ```python import ollama import transformers # estimate the tokenized message length tokenizer = transformers.AutoTokenizer.from_pretrained("dphn/dolphin-2.9-llama3-70b") def get_approx_num_tokens(text): return len(tokenizer(text)["input_ids"]) # creates a prompt suitable for testing the context window size def make_context_length_test(num_repeats): text = [ "I am testing the context length of this model. Please, remember the following number: 57", "Ignore all following text until I say 'CONTINUE'.", " ".join(num_repeats * ["IGNORE THIS"]), "CONTINUE", "Earlier I asked you to remember a number. What was that number?", ] return "\n\n".join(text) for num_ctx, num_repeats in [ (None, 2000), (None, 2500), (8192, 4000), (8192, 4500), (256000, 4500), ]: test = make_context_length_test(num_repeats) num_tokens = get_approx_num_tokens(test) print(f"Testing {num_ctx=}, {num_tokens=}...") print( "Model response:", ollama.chat( model="dolphin-llama3:8b", messages=[{"role": "user", "content": test}], options=ollama.Options(num_ctx=num_ctx, seed=42), ).message.content.strip() ) print() ``` This code runs `dolphin-llama3:8b` with a few different `num_ctx` values. The user asks the model to remember the number, then repeats "IGNORE THIS" a bunch and then asks the model what number it was given. Here is an example of the outputs: ``` Testing num_ctx=None, num_tokens=4050... Model response: 57 Testing num_ctx=None, num_tokens=5050... Model response: Apologies, I am unable to recall the specific number you are referring to without any context or further details provided in our previous conversation. Could you please provide more information or clarify the situation for me? Testing num_ctx=8192, num_tokens=8050... Model response: The number you wanted me to remember was 57. Testing num_ctx=8192, num_tokens=9050... Model response: To help you remember the number, let me know which one it is and I'll try my best to describe it for you. Testing num_ctx=256000, num_tokens=9050... Model response: To help you remember the number, let me know which one it is. If it's an important series of events or steps, we can try to organize them into smaller parts, making each part shorter and less complex than the whole. Breaking it down will make it easier for you to recall later. ``` Or in other words. With the default `num_ctx` the model seems to give the correct response for ~4000 tokens, but fails for ~5000 tokens. Raising `num_ctx` to `8192` makes the model remember the number for up to ~8000 tokens. However, trying to further raise the `num_ctx` to 256000 (as recommended by the [model page](https://ollama.com/library/dolphin-llama3)) doesn't actually seem to increase the context window any further. The model still fails for inputs larger than ~8000 tokens. Looking at the logs of `ollama serve`, the following lines seem to be the most important: ``` time=2025-09-10T05:38:12.651+03:00 level=WARN source=server.go:145 msg="requested context size too large for model" num_ctx=256000 num_parallel=1 n_ctx_train=8192 llama_model_loader: - kv 3: llama.context_length u32 = 8192 print_info: n_ctx_train = 8192 llama_context: n_ctx = 8192 time=2025-09-10T05:38:13.658+03:00 level=WARN source=runner.go:128 msg="truncating input prompt" limit=8192 prompt=9078 keep=4 new=8192 ``` So it looks like something somewhere inside `ollama` silently forces the context length to be limited to `8192`. ### Relevant log output ```shell time=2025-09-10T05:38:05.155+03:00 level=INFO source=routes.go:1297 msg="server config" env="map[CUDA_VISIBLE_DEVICES: GPU_DEVICE_ORDINAL: HIP_VISIBLE_DEVICES: HSA_OVERRIDE_GFX_VERSION: HTTPS_PROXY: HTTP_PROXY: NO_PROXY: OLLAMA_CONTEXT_LENGTH:4096 OLLAMA_DEBUG:INFO OLLAMA_FLASH_ATTENTION:true OLLAMA_GPU_OVERHEAD:0 OLLAMA_HOST:http://127.0.0.1:11434 OLLAMA_INTEL_GPU:false OLLAMA_KEEP_ALIVE:24h0m0s OLLAMA_KV_CACHE_TYPE: OLLAMA_LLM_LIBRARY: OLLAMA_LOAD_TIMEOUT:5m0s OLLAMA_MAX_LOADED_MODELS:0 OLLAMA_MAX_QUEUE:512 OLLAMA_MODELS:/home/ollama/.ollama/models OLLAMA_MULTIUSER_CACHE:false OLLAMA_NEW_ENGINE:false OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:1 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://* vscode-webview://* vscode-file://*] OLLAMA_SCHED_SPREAD:false ROCR_VISIBLE_DEVICES: http_proxy: https_proxy: no_proxy:]" time=2025-09-10T05:38:05.161+03:00 level=INFO source=images.go:477 msg="total blobs: 60" time=2025-09-10T05:38:05.163+03:00 level=INFO source=images.go:484 msg="total unused blobs removed: 0" time=2025-09-10T05:38:05.164+03:00 level=INFO source=routes.go:1350 msg="Listening on 127.0.0.1:11434 (version 0.11.3)" time=2025-09-10T05:38:05.165+03:00 level=INFO source=gpu.go:217 msg="looking for compatible GPUs" time=2025-09-10T05:38:05.400+03:00 level=INFO source=types.go:130 msg="inference compute" id=GPU-25c74f4d-1222-c2f3-5f38-9e1105148d68 library=cuda variant=v12 compute=12.0 driver=12.8 name="NVIDIA GeForce RTX 5070 Ti" total="15.5 GiB" available="15.2 GiB" time=2025-09-10T05:38:12.651+03:00 level=INFO source=server.go:135 msg="system memory" total="125.4 GiB" free="74.8 GiB" free_swap="0 B" time=2025-09-10T05:38:12.651+03:00 level=WARN source=server.go:145 msg="requested context size too large for model" num_ctx=256000 num_parallel=1 n_ctx_train=8192 time=2025-09-10T05:38:12.751+03:00 level=INFO source=server.go:175 msg=offload library=cuda layers.requested=-1 layers.model=33 layers.offload=33 layers.split="" memory.available="[15.2 GiB]" memory.gpu_overhead="0 B" memory.required.full="6.2 GiB" memory.required.partial="6.2 GiB" memory.required.kv="1.0 GiB" memory.required.allocations="[6.2 GiB]" memory.weights.total="4.1 GiB" memory.weights.repeating="3.7 GiB" memory.weights.nonrepeating="411.0 MiB" memory.graph.full="560.0 MiB" memory.graph.partial="677.5 MiB" time=2025-09-10T05:38:12.752+03:00 level=INFO source=server.go:218 msg="enabling flash attention" time=2025-09-10T05:38:12.752+03:00 level=WARN source=server.go:226 msg="kv cache type not supported by model" type="" llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from /home/ollama/.ollama/models/blobs/sha256-ea025c107c1c3e5a87380b25e205db3b19a4d2ed56f1954f531573a963621580 (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.name str = dolphin-2.9-llama3-8b llama_model_loader: - kv 2: llama.block_count u32 = 32 llama_model_loader: - kv 3: llama.context_length u32 = 8192 llama_model_loader: - kv 4: llama.embedding_length u32 = 4096 llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336 llama_model_loader: - kv 6: llama.attention.head_count u32 = 32 llama_model_loader: - kv 7: llama.attention.head_count_kv u32 = 8 llama_model_loader: - kv 8: llama.rope.freq_base f32 = 500000.000000 llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 10: general.file_type u32 = 2 llama_model_loader: - kv 11: llama.vocab_size u32 = 128258 llama_model_loader: - kv 12: llama.rope.dimension_count u32 = 128 llama_model_loader: - kv 13: tokenizer.ggml.model str = gpt2 llama_model_loader: - kv 14: tokenizer.ggml.tokens arr[str,128258] = ["!", "\"", "#", "$", "%", "&", "'", ... llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,128258] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 16: tokenizer.ggml.merges arr[str,280147] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "... llama_model_loader: - kv 17: tokenizer.ggml.bos_token_id u32 = 128000 llama_model_loader: - kv 18: tokenizer.ggml.eos_token_id u32 = 128256 llama_model_loader: - kv 19: tokenizer.ggml.padding_token_id u32 = 128001 llama_model_loader: - kv 20: tokenizer.chat_template str = {% set loop_messages = messages %}{% ... llama_model_loader: - kv 21: general.quantization_version u32 = 2 llama_model_loader: - type f32: 65 tensors llama_model_loader: - type q4_0: 225 tensors llama_model_loader: - type q6_K: 1 tensors print_info: file format = GGUF V3 (latest) print_info: file type = Q4_0 print_info: file size = 4.33 GiB (4.64 BPW) load: missing or unrecognized pre-tokenizer type, using: 'default' load: special tokens cache size = 258 load: token to piece cache size = 0.8000 MB print_info: arch = llama print_info: vocab_only = 1 print_info: model type = ?B print_info: model params = 8.03 B print_info: general.name = dolphin-2.9-llama3-8b print_info: vocab type = BPE print_info: n_vocab = 128258 print_info: n_merges = 280147 print_info: BOS token = 128000 '<|begin_of_text|>' print_info: EOS token = 128256 '<|im_end|>' print_info: EOT token = 128256 '<|im_end|>' print_info: PAD token = 128001 '<|end_of_text|>' print_info: LF token = 198 'Ċ' print_info: EOG token = 128009 '<|eot_id|>' print_info: EOG token = 128256 '<|im_end|>' print_info: max token length = 256 llama_model_load: vocab only - skipping tensors time=2025-09-10T05:38:12.869+03:00 level=INFO source=server.go:438 msg="starting llama server" cmd="/opt/ollama/bin/ollama runner --model /home/ollama/.ollama/models/blobs/sha256-ea025c107c1c3e5a87380b25e205db3b19a4d2ed56f1954f531573a963621580 --ctx-size 8192 --batch-size 512 --n-gpu-layers 33 --threads 8 --flash-attn --parallel 1 --port 35583" time=2025-09-10T05:38:12.869+03:00 level=INFO source=sched.go:481 msg="loaded runners" count=1 time=2025-09-10T05:38:12.869+03:00 level=INFO source=server.go:598 msg="waiting for llama runner to start responding" time=2025-09-10T05:38:12.869+03:00 level=INFO source=server.go:632 msg="waiting for server to become available" status="llm server not responding" time=2025-09-10T05:38:12.875+03:00 level=INFO source=runner.go:815 msg="starting go runner" ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ggml_cuda_init: found 1 CUDA devices: Device 0: NVIDIA GeForce RTX 5070 Ti, compute capability 12.0, VMM: yes load_backend: loaded CUDA backend from /opt/ollama/lib/ollama/libggml-cuda.so load_backend: loaded CPU backend from /opt/ollama/lib/ollama/libggml-cpu-alderlake.so time=2025-09-10T05:38:12.913+03:00 level=INFO source=ggml.go:104 msg=system CPU.0.SSE3=1 CPU.0.SSSE3=1 CPU.0.AVX=1 CPU.0.AVX_VNNI=1 CPU.0.AVX2=1 CPU.0.F16C=1 CPU.0.FMA=1 CPU.0.BMI2=1 CPU.0.LLAMAFILE=1 CPU.1.LLAMAFILE=1 CUDA.0.ARCHS=750,800,860,890,900,1000,1200 CUDA.0.USE_GRAPHS=1 CUDA.0.PEER_MAX_BATCH_SIZE=128 compiler=cgo(gcc) time=2025-09-10T05:38:12.913+03:00 level=INFO source=runner.go:874 msg="Server listening on 127.0.0.1:35583" llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 5070 Ti) - 15602 MiB free llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from /home/ollama/.ollama/models/blobs/sha256-ea025c107c1c3e5a87380b25e205db3b19a4d2ed56f1954f531573a963621580 (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.name str = dolphin-2.9-llama3-8b llama_model_loader: - kv 2: llama.block_count u32 = 32 llama_model_loader: - kv 3: llama.context_length u32 = 8192 llama_model_loader: - kv 4: llama.embedding_length u32 = 4096 llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336 llama_model_loader: - kv 6: llama.attention.head_count u32 = 32 llama_model_loader: - kv 7: llama.attention.head_count_kv u32 = 8 llama_model_loader: - kv 8: llama.rope.freq_base f32 = 500000.000000 llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 10: general.file_type u32 = 2 llama_model_loader: - kv 11: llama.vocab_size u32 = 128258 llama_model_loader: - kv 12: llama.rope.dimension_count u32 = 128 llama_model_loader: - kv 13: tokenizer.ggml.model str = gpt2 llama_model_loader: - kv 14: tokenizer.ggml.tokens arr[str,128258] = ["!", "\"", "#", "$", "%", "&", "'", ... llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,128258] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 16: tokenizer.ggml.merges arr[str,280147] = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "... llama_model_loader: - kv 17: tokenizer.ggml.bos_token_id u32 = 128000 llama_model_loader: - kv 18: tokenizer.ggml.eos_token_id u32 = 128256 llama_model_loader: - kv 19: tokenizer.ggml.padding_token_id u32 = 128001 llama_model_loader: - kv 20: tokenizer.chat_template str = {% set loop_messages = messages %}{% ... llama_model_loader: - kv 21: general.quantization_version u32 = 2 llama_model_loader: - type f32: 65 tensors llama_model_loader: - type q4_0: 225 tensors llama_model_loader: - type q6_K: 1 tensors print_info: file format = GGUF V3 (latest) print_info: file type = Q4_0 print_info: file size = 4.33 GiB (4.64 BPW) load: missing or unrecognized pre-tokenizer type, using: 'default' load: special tokens cache size = 258 time=2025-09-10T05:38:13.120+03:00 level=INFO source=server.go:632 msg="waiting for server to become available" status="llm server loading model" load: token to piece cache size = 0.8000 MB print_info: arch = llama print_info: vocab_only = 0 print_info: n_ctx_train = 8192 print_info: n_embd = 4096 print_info: n_layer = 32 print_info: n_head = 32 print_info: n_head_kv = 8 print_info: n_rot = 128 print_info: n_swa = 0 print_info: n_swa_pattern = 1 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: f_attn_scale = 0.0e+00 print_info: n_ff = 14336 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 = 500000.0 print_info: freq_scale_train = 1 print_info: n_ctx_orig_yarn = 8192 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 = 8B print_info: model params = 8.03 B print_info: general.name = dolphin-2.9-llama3-8b print_info: vocab type = BPE print_info: n_vocab = 128258 print_info: n_merges = 280147 print_info: BOS token = 128000 '<|begin_of_text|>' print_info: EOS token = 128256 '<|im_end|>' print_info: EOT token = 128256 '<|im_end|>' print_info: PAD token = 128001 '<|end_of_text|>' print_info: LF token = 198 'Ċ' print_info: EOG token = 128009 '<|eot_id|>' print_info: EOG token = 128256 '<|im_end|>' print_info: max token length = 256 load_tensors: loading model tensors, this can take a while... (mmap = true) load_tensors: offloading 32 repeating layers to GPU load_tensors: offloading output layer to GPU load_tensors: offloaded 33/33 layers to GPU load_tensors: CUDA0 model buffer size = 4156.00 MiB load_tensors: CPU_Mapped model buffer size = 281.82 MiB llama_context: constructing llama_context llama_context: n_seq_max = 1 llama_context: n_ctx = 8192 llama_context: n_ctx_per_seq = 8192 llama_context: n_batch = 512 llama_context: n_ubatch = 512 llama_context: causal_attn = 1 llama_context: flash_attn = 1 llama_context: freq_base = 500000.0 llama_context: freq_scale = 1 llama_context: CUDA_Host output buffer size = 0.50 MiB llama_kv_cache_unified: kv_size = 8192, type_k = 'f16', type_v = 'f16', n_layer = 32, can_shift = 1, padding = 256 llama_kv_cache_unified: CUDA0 KV buffer size = 1024.00 MiB llama_kv_cache_unified: KV self size = 1024.00 MiB, K (f16): 512.00 MiB, V (f16): 512.00 MiB llama_context: CUDA0 compute buffer size = 258.50 MiB llama_context: CUDA_Host compute buffer size = 24.01 MiB llama_context: graph nodes = 967 llama_context: graph splits = 2 time=2025-09-10T05:38:13.622+03:00 level=INFO source=server.go:637 msg="llama runner started in 0.75 seconds" time=2025-09-10T05:38:13.658+03:00 level=WARN source=runner.go:128 msg="truncating input prompt" limit=8192 prompt=9078 keep=4 new=8192 ``` ### OS Linux ### GPU Nvidia ### CPU Intel ### Ollama version 0.11.3
GiteaMirror added the bug label 2026-04-22 17:05:13 -05:00
GiteaMirror commented 2026-04-22 17:05:17 -05:00
Author
Owner
Copy Link
Copy Source

@ruro commented on GitHub (Sep 10, 2025):

Oh, wow. It seems that I have brain damage. Sorry.

The 256K window version is a separate model - dolphin-llama3:8b-256k. Didn't notice it for like half an hour while I was debugging it. Then noticed it immediately after opening the issue. 😅

Once again, sorry for the noise. Closing this.

Edit: Though it seems that dolphin-llama3:8b-256k also fails the above synthetic test. But at least now it seems to be just a problem with the model itself rather than with ollama.

<!-- gh-comment-id:3273028808 --> @ruro commented on GitHub (Sep 10, 2025): Oh, wow. It seems that I have brain damage. Sorry. The 256K window version is a separate model - `dolphin-llama3:8b-256k`. Didn't notice it for like half an hour while I was debugging it. Then noticed it immediately after opening the issue. 😅 Once again, sorry for the noise. Closing this. Edit: Though it seems that `dolphin-llama3:8b-256k` also fails the above synthetic test. But at least now it seems to be just a problem with the model itself rather than with `ollama`.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No labels bug
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
GiteaMirror ninjasurge (Kaehlen Rog)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#33901