Eval bug: --fit-target overflows on Windows, setting a target larger than 4095 MiB is not possible

### Name and Version

version: 8241 (62b8143ad)
built with Clang 19.1.5 for Windows x86_64

### Operating systems

Windows

### GGML backends

CUDA

### Hardware

RTX 5090 + RTX 4090 + RPC on a NVIDIA GB10 based machine (Asus GX10, similar to a DGX Spark)

The issue is not specific to this hardware and the CUDA / RPC backends, it can be reproduced with any GPU.

### Models

lmstudio-community/Qwen3.5-397B-A17B-GGUF:Q4_K_M

The issue is not model related, it can be reproduced with any model.

### Problem description & steps to reproduce

When using values higher than 4095 MiB for the `--fit-target` on Windows, an integer overflow occurs. As a result, you can't set a target higher than 4095 MiB on Windows.

Since the issue is `fit` related, it can't be reproduced with `-fit off`

For example, running with `--fit-target 4095,4096,4097` for my 3 backends results in this output during initialization:

```
0.02.832.376 I common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
0.03.911.357 I llama_params_fit_impl: projected memory use with initial parameters [MiB]:
0.03.911.365 I llama_params_fit_impl:   - RPC0 (gx10-bb36:50052)         : 122502 total, 164575 used, -46590 free vs. target of   4095
0.03.911.366 I llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 5090):  32606 total,  43674 used, -12881 free vs. target of      0
0.03.911.366 I llama_params_fit_impl:   - CUDA1 (NVIDIA GeForce RTX 4090):  24563 total,  30862 used,  -7958 free vs. target of      1
0.03.911.367 I llama_params_fit_impl: projected to use 239111 MiB of device memory vs. 171682 MiB of free device memory
0.03.911.367 I llama_params_fit_impl: cannot meet free memory targets on all devices, need to use 71525 MiB less in total
0.03.911.367 I llama_params_fit_impl: context size set by user to 262144 -> no change
0.04.322.483 I llama_params_fit_impl: with only dense weights in device memory there is a total surplus of 148469 MiB
0.04.804.938 I llama_params_fit_impl: filling dense-only layers back-to-front:
0.05.264.416 I llama_params_fit_impl:   - CUDA1 (NVIDIA GeForce RTX 4090): 61 layers,  18938 MiB used,   3965 MiB free
0.05.264.424 I llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 5090):  0 layers,   2840 MiB used,  27952 MiB free
0.05.264.425 I llama_params_fit_impl:   - RPC0 (gx10-bb36:50052)         :  0 layers,      0 MiB used, 117985 MiB free
0.05.264.426 I llama_params_fit_impl: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:
0.09.141.830 I llama_params_fit_impl:   - RPC0 (gx10-bb36:50052)         : 29 layers ( 1 overflowing), 113619 MiB used,   4366 MiB free
0.15.511.768 I llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 5090):  7 layers ( 1 overflowing),  29761 MiB used,   1031 MiB free
0.18.824.381 I llama_params_fit_impl:   - CUDA1 (NVIDIA GeForce RTX 4090): 25 layers (21 overflowing),  22644 MiB used,    259 MiB free
0.18.824.391 I llama_params_fit: successfully fit params to free device memory
```

As you can see, `4096` results in an actual target of `0` and  and `4097` results in an actual target of `1`.

Interrupting the process confirms that it was actually not using the correct targets for those two devices:

```
8.21.980.896 I srv    operator(): operator(): cleaning up before exit...
8.21.984.092 I llama_memory_breakdown_print: | memory breakdown [MiB]     |  total   free      self    model   context   compute    unaccounted |
8.21.984.095 I llama_memory_breakdown_print: |   - RPC0 (gx10-bb36:50052) | 122502 = 3130 + (113619 = 108831 +    3948 +     840) +        5752 |
8.21.984.096 I llama_memory_breakdown_print: |   - CUDA0 (RTX 5090)       |  32606 =  795 + ( 29761 =  25813 +    1106 +    2840) +        2050 |
8.21.984.096 I llama_memory_breakdown_print: |   - CUDA1 (RTX 4090)       |  24563 =    0 + ( 22652 =  18773 +    3370 +     509) +        1910 |
8.21.984.097 I llama_memory_breakdown_print: |   - Host                   |                   76544 =  76016 +       0 +     528                |
```

As far as I can tell, this overflow is not a security issue.

I asked Claude Opus 4.6 high reasoning to look into it, and this is what it came up with:

# `--fit-target` values >= 4096 silently overflow to 0 on Windows

## Summary

The `--fit-target` (aka `-fitt`) parameter silently overflows to 0 for values >= 4096 MiB on Windows due to a 32-bit integer overflow in the argument parsing code.

## Reproduction

On Windows (any 64-bit version):

```
llama-cli -m model.gguf -fit on -fitt 4096
```

The margin is silently set to 0 instead of 4096 MiB. Values below 4096 work correctly.

## Root Cause

In `common/arg.cpp`, the parsing code converts MiB to bytes using:

```cpp
// line 2430 (single value):
std::fill(params.fit_params_target.begin(), params.fit_params_target.end(),
    std::stoul(split_arg[0]) * 1024*1024);

// line 2434 (multi-device):
params.fit_params_target[i] = std::stoul(split_arg[i]) * 1024*1024;
```

`std::stoul()` returns `unsigned long`, which is **32-bit on Windows** (even on 64-bit systems — this is a well-known MSVC/Windows ABI property). The multiplication `4096 * 1024 * 1024 = 2^32` overflows a 32-bit unsigned integer to exactly 0. Values below 4096 produce results < 2^32 and work fine.

The storage type `std::vector<size_t>` (in `common/common.h:389`) is 64-bit and can hold the value — the overflow happens before the assignment.

## Suggested Fix

Replace `std::stoul` with `std::stoull` (returns `unsigned long long`, guaranteed 64-bit on all platforms):

```cpp
// line 2430:
std::fill(params.fit_params_target.begin(), params.fit_params_target.end(),
    std::stoull(split_arg[0]) * 1024*1024);

// line 2434:
params.fit_params_target[i] = std::stoull(split_arg[i]) * 1024*1024;
```

## Environment

- Windows 11 (64-bit), MSVC toolchain
- `unsigned long` is 32-bit on Windows per the MSVC ABI
- Issue does not reproduce on Linux/macOS where `unsigned long` is 64-bit


### First Bad Commit

_No response_

### Relevant log output

<details>
<summary>Logs</summary>


```console
C:\llama.cpp\llama-b8241-bin-win-cuda-13.1-x64>.\llama-server.exe -m C:/Users/vladn/.lmstudio/models/lmstudio-community/Qwen3.5-397B-A17B-GGUF/Qwen3.5-397B-A17B-Q4_K_M-00001-of-00007.gguf --host 0.0.0.0 -c 262144 --log-timestamps --log-prefix --rpc gx10-bb36:50052 --chat-template-kwargs "{\"enable_thinking\":false}" --no-mmap --fit-target 4095,4096,4097
ggml_cuda_init: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
  Device 1: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
load_backend: loaded CUDA backend from C:\llama.cpp\llama-b8241-bin-win-cuda-13.1-x64\ggml-cuda.dll
load_backend: loaded RPC backend from C:\llama.cpp\llama-b8241-bin-win-cuda-13.1-x64\ggml-rpc.dll
load_backend: loaded CPU backend from C:\llama.cpp\llama-b8241-bin-win-cuda-13.1-x64\ggml-cpu-alderlake.dll
0.02.831.294 I main: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true
0.02.831.309 I build: 8241 (62b8143ad) with Clang 19.1.5 for Windows x86_64
0.02.831.319 I system info: n_threads = 24, n_threads_batch = 24, total_threads = 32
0.02.831.320 I
0.02.831.369 I system_info: n_threads = 24 (n_threads_batch = 24) / 32 | CUDA : ARCHS = 750,800,860,890,1200,1210 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | BLACKWELL_NATIVE_FP4 = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
0.02.831.370 I
0.02.831.373 I Running without SSL
0.02.831.384 I init: using 31 threads for HTTP server
0.02.831.458 I start: binding port with default address family
0.02.832.318 I main: loading model
0.02.832.319 I srv    load_model: loading model 'C:/Users/vladn/.lmstudio/models/lmstudio-community/Qwen3.5-397B-A17B-GGUF/Qwen3.5-397B-A17B-Q4_K_M-00001-of-00007.gguf'
0.02.832.376 I common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
0.03.911.357 I llama_params_fit_impl: projected memory use with initial parameters [MiB]:
0.03.911.365 I llama_params_fit_impl:   - RPC0 (gx10-bb36:50052)         : 122502 total, 164575 used, -46590 free vs. target of   4095
0.03.911.366 I llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 5090):  32606 total,  43674 used, -12881 free vs. target of      0
0.03.911.366 I llama_params_fit_impl:   - CUDA1 (NVIDIA GeForce RTX 4090):  24563 total,  30862 used,  -7958 free vs. target of      1
0.03.911.367 I llama_params_fit_impl: projected to use 239111 MiB of device memory vs. 171682 MiB of free device memory
0.03.911.367 I llama_params_fit_impl: cannot meet free memory targets on all devices, need to use 71525 MiB less in total
0.03.911.367 I llama_params_fit_impl: context size set by user to 262144 -> no change
0.04.322.483 I llama_params_fit_impl: with only dense weights in device memory there is a total surplus of 148469 MiB
0.04.804.938 I llama_params_fit_impl: filling dense-only layers back-to-front:
0.05.264.416 I llama_params_fit_impl:   - CUDA1 (NVIDIA GeForce RTX 4090): 61 layers,  18938 MiB used,   3965 MiB free
0.05.264.424 I llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 5090):  0 layers,   2840 MiB used,  27952 MiB free
0.05.264.425 I llama_params_fit_impl:   - RPC0 (gx10-bb36:50052)         :  0 layers,      0 MiB used, 117985 MiB free
0.05.264.426 I llama_params_fit_impl: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:
0.09.141.830 I llama_params_fit_impl:   - RPC0 (gx10-bb36:50052)         : 29 layers ( 1 overflowing), 113619 MiB used,   4366 MiB free
0.15.511.768 I llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 5090):  7 layers ( 1 overflowing),  29761 MiB used,   1031 MiB free
0.18.824.381 I llama_params_fit_impl:   - CUDA1 (NVIDIA GeForce RTX 4090): 25 layers (21 overflowing),  22644 MiB used,    259 MiB free
0.18.824.391 I llama_params_fit: successfully fit params to free device memory
0.18.824.394 I llama_params_fit: fitting params to free memory took 0.73 seconds
0.18.827.980 I llama_model_load_from_file_impl: using device RPC0 (gx10-bb36:50052) (unknown id) - 118526 MiB free
0.18.828.015 I llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 5090) (0000:08:00.0) - 30927 MiB free
0.18.828.027 I llama_model_load_from_file_impl: using device CUDA1 (NVIDIA GeForce RTX 4090) (0000:01:00.0) - 22988 MiB free
0.18.939.729 I llama_model_loader: additional 6 GGUFs metadata loaded.
0.18.939.739 I llama_model_loader: loaded meta data with 43 key-value pairs and 1098 tensors from C:/Users/vladn/.lmstudio/models/lmstudio-community/Qwen3.5-397B-A17B-GGUF/Qwen3.5-397B-A17B-Q4_K_M-00001-of-00007.gguf (version GGUF V3 (latest))
0.18.939.796 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
0.18.939.798 I llama_model_loader: - kv   0:                       general.architecture str              = qwen35moe
0.18.939.799 I llama_model_loader: - kv   1:                               general.type str              = model
0.18.939.800 I llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 20
0.18.939.804 I llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
0.18.939.820 I llama_model_loader: - kv   4:                      general.sampling.temp f32              = 0.600000
0.18.939.821 I llama_model_loader: - kv   5:                               general.name str              = Qwen_Qwen3.5 397B A17B
0.18.939.822 I llama_model_loader: - kv   6:                           general.basename str              = Qwen_Qwen3.5
0.18.939.822 I llama_model_loader: - kv   7:                         general.size_label str              = 397B
0.18.939.823 I llama_model_loader: - kv   8:                      qwen35moe.block_count u32              = 60
0.18.939.824 I llama_model_loader: - kv   9:                   qwen35moe.context_length u32              = 262144
0.18.939.824 I llama_model_loader: - kv  10:                 qwen35moe.embedding_length u32              = 4096
0.18.939.825 I llama_model_loader: - kv  11:             qwen35moe.attention.head_count u32              = 32
0.18.939.825 I llama_model_loader: - kv  12:          qwen35moe.attention.head_count_kv u32              = 2
0.18.939.834 I llama_model_loader: - kv  13:          qwen35moe.rope.dimension_sections arr[i32,4]       = [11, 11, 10, 0]
0.18.939.836 I llama_model_loader: - kv  14:                   qwen35moe.rope.freq_base f32              = 10000000.000000
0.18.939.837 I llama_model_loader: - kv  15: qwen35moe.attention.layer_norm_rms_epsilon f32              = 0.000001
0.18.939.838 I llama_model_loader: - kv  16:                     qwen35moe.expert_count u32              = 512
0.18.939.838 I llama_model_loader: - kv  17:                qwen35moe.expert_used_count u32              = 10
0.18.939.839 I llama_model_loader: - kv  18:             qwen35moe.attention.key_length u32              = 256
0.18.939.839 I llama_model_loader: - kv  19:           qwen35moe.attention.value_length u32              = 256
0.18.939.840 I llama_model_loader: - kv  20:       qwen35moe.expert_feed_forward_length u32              = 1024
0.18.939.840 I llama_model_loader: - kv  21: qwen35moe.expert_shared_feed_forward_length u32              = 1024
0.18.939.841 I llama_model_loader: - kv  22:                  qwen35moe.ssm.conv_kernel u32              = 4
0.18.939.841 I llama_model_loader: - kv  23:                   qwen35moe.ssm.state_size u32              = 128
0.18.939.842 I llama_model_loader: - kv  24:                  qwen35moe.ssm.group_count u32              = 16
0.18.939.842 I llama_model_loader: - kv  25:               qwen35moe.ssm.time_step_rank u32              = 64
0.18.939.843 I llama_model_loader: - kv  26:                   qwen35moe.ssm.inner_size u32              = 8192
0.18.939.843 I llama_model_loader: - kv  27:          qwen35moe.full_attention_interval u32              = 4
0.18.939.844 I llama_model_loader: - kv  28:             qwen35moe.rope.dimension_count u32              = 64
0.18.939.844 I llama_model_loader: - kv  29:                       tokenizer.ggml.model str              = gpt2
0.18.939.845 I llama_model_loader: - kv  30:                         tokenizer.ggml.pre str              = qwen35
0.18.992.422 I llama_model_loader: - kv  31:                      tokenizer.ggml.tokens arr[str,248320]  = ["!", "\"", "#", "$", "%", "&", "'", ...
0.19.014.442 I llama_model_loader: - kv  32:                  tokenizer.ggml.token_type arr[i32,248320]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
0.19.087.568 I llama_model_loader: - kv  33:                      tokenizer.ggml.merges arr[str,247587]  = ["─á ─á", "─á─á ─á─á", "i n", "─á t",...
0.19.087.580 I llama_model_loader: - kv  34:                tokenizer.ggml.eos_token_id u32              = 248046
0.19.087.581 I llama_model_loader: - kv  35:            tokenizer.ggml.padding_token_id u32              = 248044
0.19.087.584 I llama_model_loader: - kv  36:               tokenizer.ggml.add_bos_token bool             = false
0.19.087.591 I llama_model_loader: - kv  37:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
0.19.087.592 I llama_model_loader: - kv  38:               general.quantization_version u32              = 2
0.19.087.593 I llama_model_loader: - kv  39:                          general.file_type u32              = 15
0.19.087.594 I llama_model_loader: - kv  40:                                   split.no u16              = 0
0.19.087.595 I llama_model_loader: - kv  41:                        split.tensors.count i32              = 1098
0.19.087.595 I llama_model_loader: - kv  42:                                split.count u16              = 7
0.19.087.596 I llama_model_loader: - type  f32:  451 tensors
0.19.087.597 I llama_model_loader: - type q4_K:  532 tensors
0.19.087.598 I llama_model_loader: - type q5_K:   45 tensors
0.19.087.598 I llama_model_loader: - type q6_K:   70 tensors
0.19.087.599 I print_info: file format = GGUF V3 (latest)
0.19.087.600 I print_info: file type   = Q4_K - Medium
0.19.087.604 I print_info: file size   = 224.06 GiB (4.86 BPW)
0.19.281.738 I load: 0 unused tokens
0.19.336.179 I load: printing all EOG tokens:
0.19.336.187 I load:   - 248044 ('<|endoftext|>')
0.19.336.187 I load:   - 248046 ('<|im_end|>')
0.19.336.188 I load:   - 248063 ('<|fim_pad|>')
0.19.336.188 I load:   - 248064 ('<|repo_name|>')
0.19.336.189 I load:   - 248065 ('<|file_sep|>')
0.19.342.104 I load: special tokens cache size = 33
0.19.408.217 I load: token to piece cache size = 1.7581 MB
0.19.408.234 I print_info: arch                  = qwen35moe
0.19.408.235 I print_info: vocab_only            = 0
0.19.408.235 I print_info: no_alloc              = 0
0.19.408.235 I print_info: n_ctx_train           = 262144
0.19.408.236 I print_info: n_embd                = 4096
0.19.408.236 I print_info: n_embd_inp            = 4096
0.19.408.237 I print_info: n_layer               = 60
0.19.408.247 I print_info: n_head                = 32
0.19.408.249 I print_info: n_head_kv             = 2
0.19.408.250 I print_info: n_rot                 = 64
0.19.408.250 I print_info: n_swa                 = 0
0.19.408.251 I print_info: is_swa_any            = 0
0.19.408.251 I print_info: n_embd_head_k         = 256
0.19.408.251 I print_info: n_embd_head_v         = 256
0.19.408.254 I print_info: n_gqa                 = 16
0.19.408.256 I print_info: n_embd_k_gqa          = 512
0.19.408.258 I print_info: n_embd_v_gqa          = 512
0.19.408.259 I print_info: f_norm_eps            = 0.0e+00
0.19.408.261 I print_info: f_norm_rms_eps        = 1.0e-06
0.19.408.261 I print_info: f_clamp_kqv           = 0.0e+00
0.19.408.261 I print_info: f_max_alibi_bias      = 0.0e+00
0.19.408.262 I print_info: f_logit_scale         = 0.0e+00
0.19.408.262 I print_info: f_attn_scale          = 0.0e+00
0.19.408.264 I print_info: n_ff                  = 0
0.19.408.265 I print_info: n_expert              = 512
0.19.408.265 I print_info: n_expert_used         = 10
0.19.408.265 I print_info: n_expert_groups       = 0
0.19.408.266 I print_info: n_group_used          = 0
0.19.408.266 I print_info: causal attn           = 1
0.19.408.266 I print_info: pooling type          = 0
0.19.408.267 I print_info: rope type             = 40
0.19.408.267 I print_info: rope scaling          = linear
0.19.408.268 I print_info: freq_base_train       = 10000000.0
0.19.408.269 I print_info: freq_scale_train      = 1
0.19.408.270 I print_info: n_ctx_orig_yarn       = 262144
0.19.408.270 I print_info: rope_yarn_log_mul     = 0.0000
0.19.408.271 I print_info: rope_finetuned        = unknown
0.19.408.271 I print_info: mrope sections        = [11, 11, 10, 0]
0.19.408.271 I print_info: ssm_d_conv            = 4
0.19.408.272 I print_info: ssm_d_inner           = 8192
0.19.408.272 I print_info: ssm_d_state           = 128
0.19.408.272 I print_info: ssm_dt_rank           = 64
0.19.408.273 I print_info: ssm_n_group           = 16
0.19.408.273 I print_info: ssm_dt_b_c_rms        = 0
0.19.408.273 I print_info: model type            = 397B.A17B
0.19.408.274 I print_info: model params          = 396.35 B
0.19.408.275 I print_info: general.name          = Qwen_Qwen3.5 397B A17B
0.19.408.276 I print_info: vocab type            = BPE
0.19.408.276 I print_info: n_vocab               = 248320
0.19.408.276 I print_info: n_merges              = 247587
0.19.408.277 I print_info: BOS token             = 11 ','
0.19.408.277 I print_info: EOS token             = 248046 '<|im_end|>'
0.19.408.278 I print_info: EOT token             = 248046 '<|im_end|>'
0.19.408.278 I print_info: PAD token             = 248044 '<|endoftext|>'
0.19.408.279 I print_info: LF token              = 198 '─è'
0.19.408.279 I print_info: FIM PRE token         = 248060 '<|fim_prefix|>'
0.19.408.280 I print_info: FIM SUF token         = 248062 '<|fim_suffix|>'
0.19.408.280 I print_info: FIM MID token         = 248061 '<|fim_middle|>'
0.19.408.280 I print_info: FIM PAD token         = 248063 '<|fim_pad|>'
0.19.408.281 I print_info: FIM REP token         = 248064 '<|repo_name|>'
0.19.408.281 I print_info: FIM SEP token         = 248065 '<|file_sep|>'
0.19.408.282 I print_info: EOG token             = 248044 '<|endoftext|>'
0.19.408.282 I print_info: EOG token             = 248046 '<|im_end|>'
0.19.408.282 I print_info: EOG token             = 248063 '<|fim_pad|>'
0.19.408.283 I print_info: EOG token             = 248064 '<|repo_name|>'
0.19.408.283 I print_info: EOG token             = 248065 '<|file_sep|>'
0.19.408.283 I print_info: max token length      = 256
0.19.408.285 I load_tensors: loading model tensors, this can take a while... (mmap = false, direct_io = false)
0.22.996.045 I load_tensors: offloading output layer to GPU
0.22.996.049 I load_tensors: offloading 59 repeating layers to GPU
0.22.996.049 I load_tensors: offloaded 61/61 layers to GPU
0.22.996.053 I load_tensors:          CPU model buffer size =   545.62 MiB
0.22.996.057 I load_tensors:        CUDA0 model buffer size = 25813.49 MiB
0.22.996.058 I load_tensors:        CUDA1 model buffer size = 18773.76 MiB
0.22.996.058 I load_tensors:          CPU model buffer size = 75470.77 MiB
0.22.996.059 I load_tensors: RPC0[gx10-bb36:50052] model buffer size = 108831.00 MiB
....................................................................................................
8.06.815.206 I common_init_result: added <|endoftext|> logit bias = -inf
8.06.815.212 I common_init_result: added <|im_end|> logit bias = -inf
8.06.815.213 I common_init_result: added <|fim_pad|> logit bias = -inf
8.06.815.214 I common_init_result: added <|repo_name|> logit bias = -inf
8.06.815.214 I common_init_result: added <|file_sep|> logit bias = -inf
8.06.822.523 I llama_context: constructing llama_context
8.06.822.544 I llama_context: n_seq_max     = 4
8.06.822.545 I llama_context: n_ctx         = 262144
8.06.822.545 I llama_context: n_ctx_seq     = 262144
8.06.822.666 I llama_context: n_batch       = 2048
8.06.822.718 I llama_context: n_ubatch      = 512
8.06.822.719 I llama_context: causal_attn   = 1
8.06.822.720 I llama_context: flash_attn    = auto
8.06.822.720 I llama_context: kv_unified    = true
8.06.822.724 I llama_context: freq_base     = 10000000.0
8.06.822.725 I llama_context: freq_scale    = 1
8.06.824.140 I llama_context:  CUDA_Host  output buffer size =     3.79 MiB
8.06.830.117 I llama_kv_cache:      CUDA0 KV buffer size =  1024.00 MiB
8.06.902.458 I llama_kv_cache:      CUDA1 KV buffer size =  3072.00 MiB
8.08.963.930 I llama_kv_cache: RPC0[gx10-bb36:50052] KV buffer size =  3584.00 MiB
8.08.983.661 I llama_kv_cache: size = 7680.00 MiB (262144 cells,  15 layers,  4/1 seqs), K (f16): 3840.00 MiB, V (f16): 3840.00 MiB
8.08.991.739 I llama_memory_recurrent:      CUDA0 RS buffer size =    82.81 MiB
8.11.224.257 I llama_memory_recurrent:      CUDA1 RS buffer size =   298.12 MiB
8.11.263.752 I llama_memory_recurrent: RPC0[gx10-bb36:50052] RS buffer size =   364.38 MiB
8.11.263.763 I llama_memory_recurrent: size =  745.31 MiB (     4 cells,  60 layers,  4 seqs), R (f32):   25.31 MiB, S (f32):  720.00 MiB
8.11.263.774 I sched_reserve: reserving ...
8.11.298.538 I sched_reserve: Flash Attention was auto, set to enabled
8.12.209.851 I sched_reserve: RPC0[gx10-bb36:50052] compute buffer size =   840.02 MiB
8.12.209.860 I sched_reserve:      CUDA0 compute buffer size =  2841.00 MiB
8.12.209.861 I sched_reserve:      CUDA1 compute buffer size =   509.00 MiB
8.12.209.862 I sched_reserve:        CPU compute buffer size =   528.02 MiB
8.12.209.864 I sched_reserve: graph nodes  = 10284 (with bs=512), 5919 (with bs=1)
8.12.209.865 I sched_reserve: graph splits = 97 (with bs=512), 52 (with bs=1)
8.12.209.866 I sched_reserve: reserve took 946.09 ms, sched copies = 1
8.12.211.447 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
8.16.302.338 I srv    load_model: initializing slots, n_slots = 4
8.19.105.359 W common_speculative_is_compat: the target context does not support partial sequence removal
8.19.137.062 W srv    load_model: speculative decoding not supported by this context
8.19.137.071 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 262144
8.19.137.734 I slot   load_model: id  1 | task -1 | new slot, n_ctx = 262144
8.19.137.743 I slot   load_model: id  2 | task -1 | new slot, n_ctx = 262144
8.19.137.747 I slot   load_model: id  3 | task -1 | new slot, n_ctx = 262144
8.19.137.844 W srv    load_model: prompt cache is enabled, size limit: 8192 MiB
8.19.137.844 W srv    load_model: use `--cache-ram 0` to disable the prompt cache
8.19.137.845 W srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
8.19.175.519 I init: chat template, example_format: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
<think>

</think>

'
8.19.194.944 I srv          init: init: chat template, thinking = 1
8.19.194.972 I main: model loaded
8.19.194.974 I main: server is listening on http://0.0.0.0:8080
8.19.194.975 I main: starting the main loop...
8.19.194.987 I srv  update_slots: all slots are idle
8.21.980.896 I srv    operator(): operator(): cleaning up before exit...
8.21.984.092 I llama_memory_breakdown_print: | memory breakdown [MiB]     |  total   free      self    model   context   compute    unaccounted |
8.21.984.095 I llama_memory_breakdown_print: |   - RPC0 (gx10-bb36:50052) | 122502 = 3130 + (113619 = 108831 +    3948 +     840) +        5752 |
8.21.984.096 I llama_memory_breakdown_print: |   - CUDA0 (RTX 5090)       |  32606 =  795 + ( 29761 =  25813 +    1106 +    2840) +        2050 |
8.21.984.096 I llama_memory_breakdown_print: |   - CUDA1 (RTX 4090)       |  24563 =    0 + ( 22652 =  18773 +    3370 +     509) +        1910 |
8.21.984.097 I llama_memory_breakdown_print: |   - Host                   |                   76544 =  76016 +       0 +     528                |
```
</details>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Eval bug: --fit-target overflows on Windows, setting a target larger than 4095 MiB is not possible #20308

Name and Version

Operating systems

GGML backends

Hardware

Models

Problem description & steps to reproduce

`--fit-target` values >= 4096 silently overflow to 0 on Windows

Summary

Reproduction

Root Cause

Suggested Fix

Environment

First Bad Commit

Relevant log output

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Eval bug: --fit-target overflows on Windows, setting a target larger than 4095 MiB is not possible #20308

Description

Name and Version

Operating systems

GGML backends

Hardware

Models

Problem description & steps to reproduce

--fit-target values >= 4096 silently overflow to 0 on Windows

Summary

Reproduction

Root Cause

Suggested Fix

Environment

First Bad Commit

Relevant log output

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions

`--fit-target` values >= 4096 silently overflow to 0 on Windows