mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-05-05 13:51:45 +00:00
Add GitHub data (#637)
This commit is contained in:
Thomas
@@ -0,0 +1,781 @@
|
||||
### 🔀 [#260](https://github.com/ikawrakow/ik_llama.cpp/pull/260) - FlashMLA-2: reduce compute buffer size (CUDA and CPU)
|
||||
|
||||
| **Author** | `ikawrakow` |
|
||||
| :--- | :--- |
|
||||
| **State** | ❌ **Closed** |
|
||||
| **Created** | 2025-03-17 |
|
||||
| **Updated** | 2025-03-18 |
|
||||
|
||||
---
|
||||
|
||||
#### Description
|
||||
|
||||
This PR
|
||||
* Implements the same compute buffer size reduction approach as PR #253 on CUDA
|
||||
* Adds the ability to control the compute buffer size for FlashMLA-2 (`-mla 2 -fa`) via the `-amb` command line option.
|
||||
* Fixes a bunch of integer overflows that show up when one starts using very long contexts (in the `perplexity` tool, and in the CUDA implementation of `GGML_OP_CONCAT`)
|
||||
|
||||
For FlashMLA-2 one computes $X = W_{kv} K$, where $K$ is the K-cache and $W_{kv}$ is the `blk.*.attk_kv_b.weight` tensor. $X$ has the shape `(n_embd_k_nope + n_embd_v) x n_kv x n_head`, where `n_kv` is the number of tokens currently in the cache, `n_head` is the number of heads, and `n_embd_k_nope, n_embd_v` are the head dimensions. For DeepSeekV3/R1/Lite `n_embd_k_nope = n_embd_v = 128`. As I don't have the ability to run DeepSeekV3/R1, I'm experimenting with DeepSeek-Lite, where `n_head = 16`, so I had not noticed how large $X$ can become (it is "just" 1 GiB for a context of 65k tokens). But `n_head = 128` for DeepSeekV3/R1, so for a context of 65k tokens $X$ becomes 8 GiB ($X$ is computed as `fp32`). When attention is computed on the GPU the cache is `fp16` (quantized cache still does not work for FlashMLA-2 on CUDA), so $X$ gets converted to `fp16` tensors $V$ and $K_{\rm nope}$, both having half the elements of $X$. As all 3 tensors need to exist simultaneously before the memory used for $X$ can be reused for other data, we end up requiring 16 GiB for these 3 tensors for a context of 65k tokens. This severely limits the maximum context length that can be processed on a GPU with limited VRAM. This PR solves the problem by slitting the attention computation into chunks. The number of chunks used is determined by the size of $X$ and the maximum attention buffer size $B_{\rm max}$ specified on the command-line via the `-amb` option (the argument following `-amb` is maximum buffer size in MiB). We have $N_{\rm step} = {\rm sizeof}(X)/B_{\rm max}$. In each step, $1/N_{\rm srep}$ of the $W_{kv}$ matrix are used, and the entire FlashMLA-2 series of operations is processed with this reduced dataset (effectively using `n_head`/ $N_{\rm step}$ attention heads). The final attention result is obtained by concatenating the results of the individual steps along the head dimension.
|
||||
|
||||
For DeepSeek-Lite I need to use a quite low `-amb` threshold of 256 MiB to even trigger the multi-step attention calculation at 65k tokens (attention is computed with 4 steps at 65k tokens, 2 steps at 32k tokens, and 1 step for 16k tokens or less). I observe a 2-3% drop in performance on the CPU and on CUDA for context of 32k tokens computed in 2 steps. I would really appreciate if someone tested this PR with DeepSeekV3/R1 and reported
|
||||
* Compute buffer size at 16k, 32k, 65k tokens using, e.g., `-mla 2 -fa -amb 1024 -fmoe`
|
||||
* Performance relative to not using `-amb 1024` (only PP performance is required, TG in FlashMLA-2 is done the same way as no FA, so does not go through this memory optimization).
|
||||
|
||||
---
|
||||
|
||||
#### 💬 Conversation
|
||||
|
||||
👤 **davidsyoung** commented the **2025-03-17** at **15:00:38**:<br>
|
||||
|
||||
First model load:
|
||||
|
||||
```
|
||||
./llama-server -m /models/DeepSeek-R1-GGUF/DeepSeek-R1-GGUF-iq4_xs__iq3_s_q8.gguf -amb 1024 -fmoe -mla 2 -fa -ts 24/24/24/24/24/24/24/24/24/24/24/24/24/24/24/24 -c 16384 -ub 1024 --n-gpu-layers 100 -ot "blk\.3\.ffn_(down|gate|up)_exps\.weight|blk\.4\.ffn_(down|gate|up)_exps\.weight|blk\.5\.ffn_(down|gate|up)_exps\.weight=CUDA0" -ot "blk\.6\.ffn_(down|gate|up)_exps\.weight|blk\.7\.ffn_(down|gate|up)_exps\.weight|blk\.8\.ffn_(down|gate|up)_exps\.weight=CUDA1" -ot "blk\.9\.ffn_(down|gate|up)_exps\.weight|blk\.10\.ffn_(down|gate|up)_exps\.weight|blk\.11\.ffn_(down|gate|up)_exps\.weight|blk\.12\.ffn_(down|gate|up)_exps\.weight=CUDA2" -ot "blk\.13\.ffn_(down|gate|up)_exps\.weight|blk\.14\.ffn_(down|gate|up)_exps\.weight|blk\.15\.ffn_(down|gate|up)_exps\.weight|blk\.16\.ffn_(down|gate|up)_exps\.weight=CUDA3" -ot "blk\.17\.ffn_(down|gate|up)_exps\.weight|blk\.18\.ffn_(down|gate|up)_exps\.weight|blk\.19\.ffn_(down|gate|up)_exps\.weight|blk\.20\.ffn_(down|gate|up)_exps\.weight=CUDA4" -ot "blk\.21\.ffn_(down|gate|up)_exps\.weight|blk\.22\.ffn_(down|gate|up)_exps\.weight|blk\.23\.ffn_(down|gate|up)_exps\.weight|blk\.24\.ffn_(down|gate|up)_exps\.weight=CUDA5" -ot "blk\.25\.ffn_(down|gate|up)_exps\.weight|blk\.26\.ffn_(down|gate|up)_exps\.weight|blk\.27\.ffn_(down|gate|up)_exps\.weight|blk\.28\.ffn_(down|gate|up)_exps\.weight=CUDA6" -ot "blk\.29\.ffn_(down|gate|up)_exps\.weight|blk\.30\.ffn_(down|gate|up)_exps\.weight|blk\.31\.ffn_(down|gate|up)_exps\.weight|blk\.32\.ffn_(down|gate|up)_exps\.weight=CUDA7" -ot "blk\.33\.ffn_(down|gate|up)_exps\.weight|blk\.34\.ffn_(down|gate|up)_exps\.weight|blk\.35\.ffn_(down|gate|up)_exps\.weight|blk\.36\.ffn_(down|gate|up)_exps\.weight=CUDA8" -ot "blk\.37\.ffn_(down|gate|up)_exps\.weight|blk\.38\.ffn_(down|gate|up)_exps\.weight|blk\.39\.ffn_(down|gate|up)_exps\.weight|blk\.40\.ffn_(down|gate|up)_exps\.weight=CUDA9" -ot "blk\.41\.ffn_(down|gate|up)_exps\.weight|blk\.42\.ffn_(down|gate|up)_exps\.weight|blk\.43\.ffn_(down|gate|up)_exps\.weight|blk\.44\.ffn_(down|gate|up)_exps\.weight=CUDA10" -ot "blk\.45\.ffn_(down|gate|up)_exps\.weight|blk\.46\.ffn_(down|gate|up)_exps\.weight|blk\.47\.ffn_(down|gate|up)_exps\.weight|blk\.48\.ffn_(down|gate|up)_exps\.weight=CUDA11" -ot "blk\.49\.ffn_(down|gate|up)_exps\.weight|blk\.50\.ffn_(down|gate|up)_exps\.weight|blk\.51\.ffn_(down|gate|up)_exps\.weight|blk\.52\.ffn_(down|gate|up)_exps\.weight=CUDA12" -ot "blk\.53\.ffn_(down|gate|up)_exps\.weight|blk\.54\.ffn_(down|gate|up)_exps\.weight|blk\.55\.ffn_(down|gate|up)_exps\.weight|blk\.56\.ffn_(down|gate|up)_exps\.weight=CUDA13" -ot "blk\.57\.ffn_(down|gate|up)_exps\.weight|blk\.58\.ffn_(down|gate|up)_exps\.weight|blk\.59\.ffn_(down|gate|up)_exps\.weight|blk\.60\.ffn_(down|gate|up)_exps\.weight=CUDA14" --seed 3704 --temp 0.5 --temp 0.5 --host 0.0.0.0 --port 8080
|
||||
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
|
||||
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
|
||||
ggml_cuda_init: found 16 CUDA devices:
|
||||
Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 1: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 2: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 3: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 4: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 5: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 6: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 7: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 8: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 9: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 10: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 11: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 12: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 13: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 14: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
Device 15: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
|
||||
INFO [ main] build info | tid="22757404872704" timestamp=1742222860 build=0 commit="unknown"
|
||||
INFO [ main] system info | tid="22757404872704" timestamp=1742222860 n_threads=64 n_threads_batch=-1 total_threads=128 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | "
|
||||
llama_model_loader: loaded meta data with 54 key-value pairs and 1147 tensors from /models/DeepSeek-R1-GGUF/DeepSeek-R1-GGUF-iq4_xs__iq3_s_q8.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 = deepseek2
|
||||
llama_model_loader: - kv 1: general.type str = model
|
||||
llama_model_loader: - kv 2: general.name str = unsloth_DeepSeek R1 BF16
|
||||
llama_model_loader: - kv 3: general.size_label str = 256x21B
|
||||
llama_model_loader: - kv 4: general.license str = mit
|
||||
llama_model_loader: - kv 5: general.base_model.count u32 = 1
|
||||
llama_model_loader: - kv 6: general.base_model.0.name str = DeepSeek R1
|
||||
llama_model_loader: - kv 7: general.base_model.0.organization str = Deepseek Ai
|
||||
llama_model_loader: - kv 8: general.base_model.0.repo_url str = https://huggingface.co/deepseek-ai/De...
|
||||
llama_model_loader: - kv 9: general.tags arr[str,3] = ["deepseek", "unsloth", "transformers"]
|
||||
llama_model_loader: - kv 10: general.languages arr[str,1] = ["en"]
|
||||
llama_model_loader: - kv 11: deepseek2.block_count u32 = 61
|
||||
llama_model_loader: - kv 12: deepseek2.context_length u32 = 163840
|
||||
llama_model_loader: - kv 13: deepseek2.embedding_length u32 = 7168
|
||||
llama_model_loader: - kv 14: deepseek2.feed_forward_length u32 = 18432
|
||||
llama_model_loader: - kv 15: deepseek2.attention.head_count u32 = 128
|
||||
llama_model_loader: - kv 16: deepseek2.attention.head_count_kv u32 = 128
|
||||
llama_model_loader: - kv 17: deepseek2.rope.freq_base f32 = 10000.000000
|
||||
llama_model_loader: - kv 18: deepseek2.attention.layer_norm_rms_epsilon f32 = 0.000001
|
||||
llama_model_loader: - kv 19: deepseek2.expert_used_count u32 = 8
|
||||
llama_model_loader: - kv 20: general.file_type u32 = 7
|
||||
llama_model_loader: - kv 21: deepseek2.leading_dense_block_count u32 = 3
|
||||
llama_model_loader: - kv 22: deepseek2.vocab_size u32 = 129280
|
||||
llama_model_loader: - kv 23: deepseek2.attention.q_lora_rank u32 = 1536
|
||||
llama_model_loader: - kv 24: deepseek2.attention.kv_lora_rank u32 = 512
|
||||
llama_model_loader: - kv 25: deepseek2.attention.key_length u32 = 192
|
||||
llama_model_loader: - kv 26: deepseek2.attention.value_length u32 = 128
|
||||
llama_model_loader: - kv 27: deepseek2.expert_feed_forward_length u32 = 2048
|
||||
llama_model_loader: - kv 28: deepseek2.expert_count u32 = 256
|
||||
llama_model_loader: - kv 29: deepseek2.expert_shared_count u32 = 1
|
||||
llama_model_loader: - kv 30: deepseek2.expert_weights_scale f32 = 2.500000
|
||||
llama_model_loader: - kv 31: deepseek2.expert_weights_norm bool = true
|
||||
llama_model_loader: - kv 32: deepseek2.expert_gating_func u32 = 2
|
||||
llama_model_loader: - kv 33: deepseek2.rope.dimension_count u32 = 64
|
||||
llama_model_loader: - kv 34: deepseek2.rope.scaling.type str = yarn
|
||||
llama_model_loader: - kv 35: deepseek2.rope.scaling.factor f32 = 40.000000
|
||||
llama_model_loader: - kv 36: deepseek2.rope.scaling.original_context_length u32 = 4096
|
||||
llama_model_loader: - kv 37: deepseek2.rope.scaling.yarn_log_multiplier f32 = 0.100000
|
||||
llama_model_loader: - kv 38: tokenizer.ggml.model str = gpt2
|
||||
llama_model_loader: - kv 39: tokenizer.ggml.pre str = deepseek-v3
|
||||
llama_model_loader: - kv 40: tokenizer.ggml.tokens arr[str,129280] = ["<｜begin▁of▁sentence｜>", "<?...
|
||||
llama_model_loader: - kv 41: tokenizer.ggml.token_type arr[i32,129280] = [3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
|
||||
llama_model_loader: - kv 42: tokenizer.ggml.merges arr[str,127741] = ["Ġ t", "Ġ a", "i n", "Ġ Ġ", "h e...
|
||||
llama_model_loader: - kv 43: tokenizer.ggml.bos_token_id u32 = 0
|
||||
llama_model_loader: - kv 44: tokenizer.ggml.eos_token_id u32 = 1
|
||||
llama_model_loader: - kv 45: tokenizer.ggml.padding_token_id u32 = 128815
|
||||
llama_model_loader: - kv 46: tokenizer.ggml.add_bos_token bool = true
|
||||
llama_model_loader: - kv 47: tokenizer.ggml.add_eos_token bool = false
|
||||
llama_model_loader: - kv 48: tokenizer.chat_template str = {% if not add_generation_prompt is de...
|
||||
llama_model_loader: - kv 49: general.quantization_version u32 = 2
|
||||
llama_model_loader: - kv 50: quantize.imatrix.file str = /models/deepseek-config/imatrix.dat
|
||||
llama_model_loader: - kv 51: quantize.imatrix.dataset str = imatrix-training-full-3
|
||||
llama_model_loader: - kv 52: quantize.imatrix.entries_count i32 = 720
|
||||
llama_model_loader: - kv 53: quantize.imatrix.chunks_count i32 = 315
|
||||
llama_model_loader: - type f32: 361 tensors
|
||||
llama_model_loader: - type q8_0: 438 tensors
|
||||
llama_model_loader: - type q5_K: 180 tensors
|
||||
llama_model_loader: - type iq3_s: 104 tensors
|
||||
llama_model_loader: - type iq4_xs: 64 tensors
|
||||
llm_load_vocab: special tokens cache size = 819
|
||||
llm_load_vocab: token to piece cache size = 0.8223 MB
|
||||
llm_load_print_meta: format = GGUF V3 (latest)
|
||||
llm_load_print_meta: arch = deepseek2
|
||||
llm_load_print_meta: vocab type = BPE
|
||||
llm_load_print_meta: n_vocab = 129280
|
||||
llm_load_print_meta: n_merges = 127741
|
||||
llm_load_print_meta: vocab_only = 0
|
||||
llm_load_print_meta: n_ctx_train = 163840
|
||||
llm_load_print_meta: n_embd = 7168
|
||||
llm_load_print_meta: n_layer = 61
|
||||
llm_load_print_meta: n_head = 128
|
||||
llm_load_print_meta: n_head_kv = 128
|
||||
llm_load_print_meta: n_rot = 64
|
||||
llm_load_print_meta: n_swa = 0
|
||||
llm_load_print_meta: n_embd_head_k = 192
|
||||
llm_load_print_meta: n_embd_head_v = 128
|
||||
llm_load_print_meta: n_gqa = 1
|
||||
llm_load_print_meta: n_embd_k_gqa = 24576
|
||||
llm_load_print_meta: n_embd_v_gqa = 16384
|
||||
llm_load_print_meta: f_norm_eps = 0.0e+00
|
||||
llm_load_print_meta: f_norm_rms_eps = 1.0e-06
|
||||
llm_load_print_meta: f_clamp_kqv = 0.0e+00
|
||||
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
|
||||
llm_load_print_meta: f_logit_scale = 0.0e+00
|
||||
llm_load_print_meta: n_ff = 18432
|
||||
llm_load_print_meta: n_expert = 256
|
||||
llm_load_print_meta: n_expert_used = 8
|
||||
llm_load_print_meta: causal attn = 1
|
||||
llm_load_print_meta: pooling type = 0
|
||||
llm_load_print_meta: rope type = 0
|
||||
llm_load_print_meta: rope scaling = yarn
|
||||
llm_load_print_meta: freq_base_train = 10000.0
|
||||
llm_load_print_meta: freq_scale_train = 0.025
|
||||
llm_load_print_meta: n_ctx_orig_yarn = 4096
|
||||
llm_load_print_meta: rope_finetuned = unknown
|
||||
llm_load_print_meta: ssm_d_conv = 0
|
||||
llm_load_print_meta: ssm_d_inner = 0
|
||||
llm_load_print_meta: ssm_d_state = 0
|
||||
llm_load_print_meta: ssm_dt_rank = 0
|
||||
llm_load_print_meta: model type = 671B
|
||||
llm_load_print_meta: model ftype = Q8_0
|
||||
llm_load_print_meta: model params = 672.050 B
|
||||
llm_load_print_meta: model size = 307.197 GiB (3.926 BPW)
|
||||
llm_load_print_meta: repeating layers = 305.363 GiB (3.914 BPW, 670.196 B parameters)
|
||||
llm_load_print_meta: general.name = unsloth_DeepSeek R1 BF16
|
||||
llm_load_print_meta: BOS token = 0 '<｜begin▁of▁sentence｜>'
|
||||
llm_load_print_meta: EOS token = 1 '<｜end▁of▁sentence｜>'
|
||||
llm_load_print_meta: PAD token = 128815 '<｜PAD▁TOKEN｜>'
|
||||
llm_load_print_meta: LF token = 131 'Ä'
|
||||
llm_load_print_meta: max token length = 256
|
||||
llm_load_print_meta: n_layer_dense_lead = 3
|
||||
llm_load_print_meta: n_lora_q = 1536
|
||||
llm_load_print_meta: n_lora_kv = 512
|
||||
llm_load_print_meta: n_ff_exp = 2048
|
||||
llm_load_print_meta: n_expert_shared = 1
|
||||
llm_load_print_meta: expert_weights_scale = 2.5
|
||||
llm_load_print_meta: expert_weights_norm = 1
|
||||
llm_load_print_meta: expert_gating_func = sigmoid
|
||||
llm_load_print_meta: rope_yarn_log_mul = 0.1000
|
||||
llm_load_tensors: ggml ctx size = 7.94 MiB
|
||||
Tensor blk.3.ffn_gate_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.3.ffn_down_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.3.ffn_up_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.4.ffn_gate_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.4.ffn_down_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.4.ffn_up_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.5.ffn_gate_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.5.ffn_down_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.5.ffn_up_exps.weight buffer type overriden to CUDA0
|
||||
Tensor blk.6.ffn_gate_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.6.ffn_down_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.6.ffn_up_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.7.ffn_gate_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.7.ffn_down_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.7.ffn_up_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.8.ffn_gate_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.8.ffn_down_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.8.ffn_up_exps.weight buffer type overriden to CUDA1
|
||||
Tensor blk.9.ffn_gate_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.9.ffn_down_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.9.ffn_up_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.10.ffn_gate_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.10.ffn_down_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.10.ffn_up_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.11.ffn_gate_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.11.ffn_down_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.11.ffn_up_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.12.ffn_gate_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.12.ffn_down_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.12.ffn_up_exps.weight buffer type overriden to CUDA2
|
||||
Tensor blk.13.ffn_gate_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.13.ffn_down_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.13.ffn_up_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.14.ffn_gate_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.14.ffn_down_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.14.ffn_up_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.15.ffn_gate_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.15.ffn_down_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.15.ffn_up_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.16.ffn_gate_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.16.ffn_down_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.16.ffn_up_exps.weight buffer type overriden to CUDA3
|
||||
Tensor blk.17.ffn_gate_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.17.ffn_down_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.17.ffn_up_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.18.ffn_gate_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.18.ffn_down_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.18.ffn_up_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.19.ffn_gate_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.19.ffn_down_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.19.ffn_up_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.20.ffn_gate_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.20.ffn_down_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.20.ffn_up_exps.weight buffer type overriden to CUDA4
|
||||
Tensor blk.21.ffn_gate_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.21.ffn_down_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.21.ffn_up_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.22.ffn_gate_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.22.ffn_down_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.22.ffn_up_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.23.ffn_gate_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.23.ffn_down_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.23.ffn_up_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.24.ffn_gate_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.24.ffn_down_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.24.ffn_up_exps.weight buffer type overriden to CUDA5
|
||||
Tensor blk.25.ffn_gate_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.25.ffn_down_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.25.ffn_up_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.26.ffn_gate_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.26.ffn_down_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.26.ffn_up_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.27.ffn_gate_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.27.ffn_down_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.27.ffn_up_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.28.ffn_gate_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.28.ffn_down_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.28.ffn_up_exps.weight buffer type overriden to CUDA6
|
||||
Tensor blk.29.ffn_gate_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.29.ffn_down_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.29.ffn_up_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.30.ffn_gate_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.30.ffn_down_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.30.ffn_up_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.31.ffn_gate_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.31.ffn_down_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.31.ffn_up_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.32.ffn_gate_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.32.ffn_down_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.32.ffn_up_exps.weight buffer type overriden to CUDA7
|
||||
Tensor blk.33.ffn_gate_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.33.ffn_down_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.33.ffn_up_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.34.ffn_gate_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.34.ffn_down_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.34.ffn_up_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.35.ffn_gate_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.35.ffn_down_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.35.ffn_up_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.36.ffn_gate_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.36.ffn_down_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.36.ffn_up_exps.weight buffer type overriden to CUDA8
|
||||
Tensor blk.37.ffn_gate_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.37.ffn_down_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.37.ffn_up_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.38.ffn_gate_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.38.ffn_down_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.38.ffn_up_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.39.ffn_gate_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.39.ffn_down_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.39.ffn_up_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.40.ffn_gate_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.40.ffn_down_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.40.ffn_up_exps.weight buffer type overriden to CUDA9
|
||||
Tensor blk.41.ffn_gate_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.41.ffn_down_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.41.ffn_up_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.42.ffn_gate_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.42.ffn_down_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.42.ffn_up_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.43.ffn_gate_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.43.ffn_down_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.43.ffn_up_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.44.ffn_gate_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.44.ffn_down_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.44.ffn_up_exps.weight buffer type overriden to CUDA10
|
||||
Tensor blk.45.ffn_gate_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.45.ffn_down_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.45.ffn_up_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.46.ffn_gate_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.46.ffn_down_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.46.ffn_up_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.47.ffn_gate_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.47.ffn_down_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.47.ffn_up_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.48.ffn_gate_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.48.ffn_down_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.48.ffn_up_exps.weight buffer type overriden to CUDA11
|
||||
Tensor blk.49.ffn_gate_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.49.ffn_down_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.49.ffn_up_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.50.ffn_gate_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.50.ffn_down_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.50.ffn_up_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.51.ffn_gate_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.51.ffn_down_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.51.ffn_up_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.52.ffn_gate_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.52.ffn_down_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.52.ffn_up_exps.weight buffer type overriden to CUDA12
|
||||
Tensor blk.53.ffn_gate_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.53.ffn_down_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.53.ffn_up_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.54.ffn_gate_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.54.ffn_down_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.54.ffn_up_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.55.ffn_gate_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.55.ffn_down_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.55.ffn_up_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.56.ffn_gate_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.56.ffn_down_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.56.ffn_up_exps.weight buffer type overriden to CUDA13
|
||||
Tensor blk.57.ffn_gate_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.57.ffn_down_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.57.ffn_up_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.58.ffn_gate_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.58.ffn_down_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.58.ffn_up_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.59.ffn_gate_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.59.ffn_down_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.59.ffn_up_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.60.ffn_gate_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.60.ffn_down_exps.weight buffer type overriden to CUDA14
|
||||
Tensor blk.60.ffn_up_exps.weight buffer type overriden to CUDA14
|
||||
llm_load_tensors: offloading 61 repeating layers to GPU
|
||||
llm_load_tensors: offloading non-repeating layers to GPU
|
||||
llm_load_tensors: offloaded 62/62 layers to GPU
|
||||
llm_load_tensors: CPU buffer size = 938.98 MiB
|
||||
llm_load_tensors: CUDA0 buffer size = 20883.36 MiB
|
||||
llm_load_tensors: CUDA1 buffer size = 19786.12 MiB
|
||||
llm_load_tensors: CUDA2 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA3 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA4 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA5 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA6 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA7 buffer size = 20663.59 MiB
|
||||
llm_load_tensors: CUDA8 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA9 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA10 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA11 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA12 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA13 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA14 buffer size = 20906.12 MiB
|
||||
llm_load_tensors: CUDA15 buffer size = 1424.07 MiB
|
||||
....................................................................................................
|
||||
llama_new_context_with_model: n_ctx = 16384
|
||||
llama_new_context_with_model: n_batch = 2048
|
||||
llama_new_context_with_model: n_ubatch = 1024
|
||||
llama_new_context_with_model: flash_attn = 1
|
||||
llama_new_context_with_model: mla_attn = 2
|
||||
llama_new_context_with_model: attn_max_b = 1024
|
||||
llama_new_context_with_model: fused_moe = 1
|
||||
llama_new_context_with_model: ser = -1, 0
|
||||
llama_new_context_with_model: freq_base = 10000.0
|
||||
llama_new_context_with_model: freq_scale = 0.025
|
||||
llama_kv_cache_init: layer 0: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 1: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 2: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 3: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 4: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 5: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 6: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 7: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 8: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 9: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 10: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 11: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 12: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 13: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 14: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 15: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 16: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 17: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 18: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 19: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 20: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 21: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 22: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 23: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 24: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 25: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 26: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 27: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 28: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 29: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 30: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 31: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 32: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 33: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 34: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 35: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 36: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 37: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 38: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 39: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 40: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 41: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 42: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 43: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 44: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 45: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 46: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 47: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 48: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 49: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 50: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 51: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 52: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 53: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 54: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 55: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 56: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 57: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 58: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 59: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: layer 60: n_embd_head_qk_rope = 64, kv_lora_rank = 512
|
||||
llama_kv_cache_init: CUDA0 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA1 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA2 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA3 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA4 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA5 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA6 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA7 KV buffer size = 54.00 MiB
|
||||
llama_kv_cache_init: CUDA8 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA9 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA10 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA11 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA12 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA13 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA14 KV buffer size = 72.00 MiB
|
||||
llama_kv_cache_init: CUDA15 KV buffer size = 36.00 MiB
|
||||
llama_new_context_with_model: KV self size = 1098.00 MiB, c^KV (f16): 1098.00 MiB, kv^T: not used
|
||||
llama_new_context_with_model: CUDA_Host output buffer size = 0.99 MiB
|
||||
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
|
||||
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 3480.01 MiB on device 0: cudaMalloc failed: out of memory
|
||||
ggml_gallocr_reserve_n: failed to allocate CUDA0 buffer of size 3649053696
|
||||
llama_new_context_with_model: failed to allocate compute buffers
|
||||
llama_init_from_gpt_params: error: failed to create context with model '/models/DeepSeek-R1-GGUF/DeepSeek-R1-GGUF-iq4_xs__iq3_s_q8.gguf'
|
||||
ERR [ load_model] unable to load model | tid="22757404872704" timestamp=1742223553 model="/models/DeepSeek-R1-GGUF/DeepSeek-R1-GGUF-iq4_xs__iq3_s_q8.gguf"
|
||||
Segmentation fault
|
||||
root@7e406a084738:/app/build/bin#
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-17** at **15:06:08**:<br>
|
||||
|
||||
> Those fixes for perplexity, do you believe that was related to NaN's in IX_K quants?
|
||||
|
||||
No. It is an integer overflow. The logic location in the array of logits was computed with 32-bit integers. As there are ~128k entries in the vocabulary, the integer multiplication `i * n_vocab` overflows for `i >= 16384`. You were computing PPL for contexts of 2048 or 512, so no issue there (`i < 2048`). The NaNs really are due to `fp16` arithmetic for the MoE matrix multiplications when using `IQ4_K` or `IQ4_KSS`. Apparently in the `llama.cpp` world it is well known that one cannot use the `fp16` DeepSeek models because one gets NaNs.
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-17** at **15:10:29**:<br>
|
||||
|
||||
> Segfault with `-c 16384 -amb 1024 -fmoe -mla 2 -fa`
|
||||
|
||||
It fails to allocate `3480 MiB`, so I guess there isn't enough VRAM? Try with `-amb 512` then.
|
||||
|
||||
---
|
||||
|
||||
👤 **ubergarm** commented the **2025-03-17** at **15:40:37**:<br>
|
||||
|
||||
I'll take a quick stab at it too given using a simple 1x RTX A6000 48GB GPU configuration.
|
||||
|
||||
|
||||
#### Update
|
||||
```bash
|
||||
$ git checkout ik/flash_mla2_cuda_no_f32
|
||||
$ git rev-parse --short HEAD
|
||||
b147e31f
|
||||
$ cmake -B ./build -DGGML_CUDA=ON -DGGML_BLAS=OFF
|
||||
$ cmake --build ./build --config Release -j $(nproc)
|
||||
$ ./build/bin/llama-server --version
|
||||
version: 3601 (b147e31f)
|
||||
built with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu
|
||||
```
|
||||
|
||||
> Compute buffer size at 16k, 32k, 65k tokens using, e.g., -mla 2 -fa -amb 1024 -fmoe
|
||||
|
||||
#### Basic Command
|
||||
```bash
|
||||
CUDA_VISIBLE_DEVICES="0," \
|
||||
./build/bin/llama-server \
|
||||
--alias ubergarm/DeepSeek-R1-Q2_K_R4 \
|
||||
--model /mnt/raid/models/ubergarm/DeepSeek-R1-GGUF/DeepSeek-R1-GGUF-Q2_K_R4.gguf \
|
||||
--ctx-size 16384 \
|
||||
-ctk f16 -ctv f16 \
|
||||
-mla 2 -fa \
|
||||
-amb 1024 \
|
||||
-fmoe \
|
||||
--n-gpu-layers 63 \
|
||||
--override-tensor exps=CPU \
|
||||
--parallel 1 \
|
||||
--threads 24 \
|
||||
--host 127.0.0.1 \
|
||||
--port 8080
|
||||
```
|
||||
|
||||
#### Results
|
||||
|
||||
* 16k = TODO
|
||||
* 32k = TODO
|
||||
* 64k = TODO
|
||||
|
||||
> Performance relative to not using -amb 1024 (only PP performance is required, TG in FlashMLA-2 is done the same way as no FA, so does not go through this memory optimization).
|
||||
|
||||
#### llama-bench
|
||||
```bash
|
||||
echo TODO
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-17** at **16:04:33**:<br>
|
||||
|
||||
So, this looks quite a bit better than the main branch. It would seem that a single 24 GB GPU could handle the non-expert tensors and up to 32k context?
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-17** at **16:33:42**:<br>
|
||||
|
||||
> Oh, I thought I noticed it was reporting less with -ctk q8_0 -ctv q8_0, I'll do a quick check and update this TODO here and confirm.
|
||||
|
||||
I haven't put a guard against using quantized cache for `mla = 2`, so it will happily initialize (and report buffer sizes), but then it will terminate when it arrives at the op that is not supported on CUDA for quantized data.
|
||||
|
||||
Based on the performance values @ubergarm posted, there doesn't seem to be any major performance impact, even with `-amb 128`. What is the compute buffer size for `-amb 128`?
|
||||
|
||||
---
|
||||
|
||||
👤 **ubergarm** commented the **2025-03-17** at **16:47:48**:<br>
|
||||
|
||||
> What is the compute buffer size for -amb 128
|
||||
|
||||
The relevant part of the above table for this specific question:
|
||||
|
||||
| commit | ctx-size | amb | ctk/ctv | CUDA0 KV buffer | CUDA0 compute buffer | nvidia-smi |
|
||||
| --- | --- | --- | --- | --- | --- | --- |
|
||||
| branch/sha | size | quant | MiB | MiB | MiB |
|
||||
| `flash_mla2_@b147e31f` | 32768 | 1024 | f16 | 2196 | 3790 | 24010 |
|
||||
| `flash_mla2_@b147e31f` | 32768 | 128 | f16 | 2196 | 2817 | 23036 |
|
||||
|
||||
---
|
||||
|
||||
👤 **davidsyoung** commented the **2025-03-17** at **16:59:59**:<br>
|
||||
|
||||
Sorry for delay here. As model loading takes quite a long amount of time on 16 GPUs, and I'm near to the limit there's been some OOMs (my own fault nothing to do with PR), I've been quite slow to come back.
|
||||
|
||||
From what I can see so far, there is zero notable difference with performance of `-amb 256` vs `-amb 512`. I would imagine that this will continue to a point. I'll test lower and create a comparison here when done.
|
||||
|
||||
TODO
|
||||
|
||||
---
|
||||
|
||||
👤 **ubergarm** commented the **2025-03-17** at **17:25:23**:<br>
|
||||
|
||||
@davidsyoung
|
||||
|
||||
> I would imagine that this will continue to a point. I'll test lower and create a comparison here when done.
|
||||
|
||||
Yeah, please double check me, but I updated my chart and command above which suggests going down to `-amb 2` is about the limit with `-amb 1` having only slightly slower pp!????
|
||||
|
||||
Curious if you have similar outcome across all your GPUs!
|
||||
|
||||
---
|
||||
|
||||
👤 **davidsyoung** commented the **2025-03-17** at **17:39:21**:<br>
|
||||
|
||||
> @davidsyoung
|
||||
>
|
||||
> > I would imagine that this will continue to a point. I'll test lower and create a comparison here when done.
|
||||
>
|
||||
> Yeah, please double check me, but I updated my chart and command above which suggests going down to `-amb 2` is about the limit with `-amb 1` having only slightly slower pp!????
|
||||
>
|
||||
> Curious if you have similar outcome across all your GPUs!
|
||||
|
||||
Interestingly I got an error for `-amb 32` when trying to maximise context length:
|
||||
```
|
||||
llama_new_context_with_model: KV self size = 1647.00 MiB, c^KV (f16): 1647.00 MiB, kv^T: not used
|
||||
llama_new_context_with_model: CUDA_Host output buffer size = 0.99 MiB
|
||||
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
|
||||
ggml_new_object: not enough space in the context's memory pool (needed 26231472, available 26231136)
|
||||
Segmentation fault
|
||||
```
|
||||
|
||||
Haven't seen that error before!
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-17** at **17:55:57**:<br>
|
||||
|
||||
Sorry, I wasn't clear enough with my request. The PP test should be done with `-p 16384` (or whatever context we are looking at). With `-p 512`, `llama-bench` will set the context to 512, so the required buffer to compute FlashMLA-2 will be quite small - `256 x 128 x 512 x 4 = 64 MiB`, so there will be more than one step only for `-amb 32` or lower. With `-amb 2` it will take 32 steps, so it will be processing 4 heads at a time. At `-amb 1` it will be 64 steps, so 2 heads per step. I find it quite surprising that we do not see performance degradation down to so many steps.
|
||||
|
||||
> Also, you should test setting -ub 1024, you should see a big difference in PP performance compared to default of -ub 512 I believe.
|
||||
|
||||
This is only relevant of the MoE experts are computed on CUDA. When the MoE part runs on the CPU the default `u_batch` size of 512 tends to give the best PP performance.
|
||||
|
||||
---
|
||||
|
||||
👤 **ubergarm** commented the **2025-03-17** at **18:00:55**:<br>
|
||||
|
||||
@davidsyoung
|
||||
|
||||
> pipeline parallelism enabled (n_copies=4)
|
||||
Hrmm, I've seen some chatter about `-DGGML_SCHED_MAX_COPIES=4` before (default). Some folks were setting it to 1. Not sure why (maybe CUDA graphs?) and that was on vanilla llama.cpp so may not apply anymore.
|
||||
|
||||
I was kinda surprised that you were offloading shared experts onto GPUs with your config given that doesn't work on ktransformers yet in my own testing an in their documentation:
|
||||
|
||||
> Note：Currently, executing experts on the GPU will conflict with CUDA Graph. Without CUDA Graph, there will be a significant slowdown. Therefore, unless you have a substantial amount of VRAM (placing a single layer of experts for DeepSeek-V3/R1 on the GPU requires at least 5.6GB of VRAM), we do not recommend enabling this feature. We are actively working on optimization. Note KExpertsTorch is untested.
|
||||
|
||||
@ikawrakow
|
||||
|
||||
> The PP test should be done with -p 16384
|
||||
|
||||
I'll set that up and post the results here soon.
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-17** at **18:09:14**:<br>
|
||||
|
||||
> I was kinda surprised that you were offloading shared experts onto GPUs with your config given that doesn't work on ktransformers yet in my own testing an in their documentation:
|
||||
|
||||
@davidsyoung has 16 x 3090's, so the entire model is run on the GPU's. CUDA graphs get disabled for MoE models (also true on mainline `llama.cpp`). Disabled CUDA graphs leading to a significant hit in performance is a myth. There is no effect for PP, and at most a few percent (< 5%, IIRC) for TG. The 16 x 3090 configuration gives ~350 t/s for PP and ~17 t/s for TG.
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-17** at **18:25:34**:<br>
|
||||
|
||||
> Interestingly I got an error for -amb 32 when trying to maximise context length:
|
||||
> Haven't seen that error before!
|
||||
|
||||
Neither have I. It means that the back-end is miscalculating the required compute buffer size somehow. Not sure what to do about that.
|
||||
|
||||
---
|
||||
|
||||
👤 **ubergarm** commented the **2025-03-17** at **19:29:15**:<br>
|
||||
|
||||
I increased `-p 16384` and set `-r 2` repetitions down from default of 5 for a quick check but it crashed before finishing with error shown below.
|
||||
|
||||
```bash
|
||||
CUDA_VISIBLE_DEVICES="0," \
|
||||
./build/bin/llama-bench \
|
||||
--model /mnt/raid/models/ubergarm/DeepSeek-R1-GGUF/DeepSeek-R1-GGUF-Q2_K_R4.gguf \
|
||||
-ctk f16 -ctv f16 \
|
||||
-mla 2 -fa 1 \
|
||||
-amb 1024,128,16,8,4,2,1 \
|
||||
-p 16384,8192 \
|
||||
-n 0 \
|
||||
-fmoe 1 \
|
||||
-r 2 \
|
||||
--n-gpu-layers 63 \
|
||||
--override-tensor exps=CPU \
|
||||
--threads 24
|
||||
|
||||
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 RTX A6000, compute capability 8.6, VMM: yes
|
||||
```
|
||||
|
||||
| model | size | params | backend | ngl | fa | mla | amb | fmoe | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --: | ----: | ---: | ------------: | ---------------: |
|
||||
| deepseek2 671B Q2_K_R4 | 238.69 GiB | 672.05 B | CUDA | 63 | 1 | 2 | 1024 | 1 | pp16384 | 84.16 ± 2.14 |
|
||||
| deepseek2 671B Q2_K_R4 | 238.69 GiB | 672.05 B | CUDA | 63 | 1 | 2 | 1024 | 1 | pp8192 | 97.67 ± 1.21 |
|
||||
| deepseek2 671B Q2_K_R4 | 238.69 GiB | 672.05 B | CUDA | 63 | 1 | 2 | 128 | 1 | pp16384 | 82.59 ± 2.70 |
|
||||
| deepseek2 671B Q2_K_R4 | 238.69 GiB | 672.05 B | CUDA | 63 | 1 | 2 | 128 | 1 | pp8192 | 96.21 ± 1.67 |
|
||||
|
||||
```
|
||||
ggml_new_object: not enough space in the context's memory pool (needed 26231472, available 26231136)
|
||||
./myscripts/benchmark.sh: line 24: 2286044 Segmentation fault (core dumped) CUDA_VISIBLE_DEVICES="0," ./build/bin/llama-bench --model /mnt/raid/models/ubergarm/DeepSeek-R1-GGUF/DeepSeek-R1-GGUF-Q2_K_R4.gguf -ctk f16 -ctv f16 -mla 2 -fa 1 -amb 1024,128,16,8,4,2,1 -p 16384,8192 -n 0 -fmoe 1 -r 2 --n-gpu-layers 63 --override-tensor exps=CPU --threads 24
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
👤 **davidsyoung** commented the **2025-03-17** at **23:37:51**:<br>
|
||||
|
||||
So compute buffers are massively improved. I don't have apples for apples comparison as I went down a rabbit hole after realising I could turn off pipeline parallel and it would also give me more VRAM back (thanks @ubergarm!). But it is massively improved.
|
||||
|
||||
Had some issues going below `-amb 32` as well with bench, but got some data:
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mla | amb | fmoe | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --: | ----: | ---: | ------------: | ---------------: |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 512 | 1 | pp16384 | 235.62 ± 3.94 |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 512 | 1 | pp8192 | 293.09 ± 0.42 |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 128 | 1 | pp16384 | 231.66 ± 0.22 |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 128 | 1 | pp8192 | 289.73 ± 0.71 |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 64 | 1 | pp16384 | 224.48 ± 0.07 |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 64 | 1 | pp8192 | 283.72 ± 0.43 |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 32 | 1 | pp16384 | 215.12 ± 0.05 |
|
||||
| deepseek2 671B Q8_0 | 307.20 GiB | 672.05 B | CUDA | 63 | 2048 | 1 | 2 | 32 | 1 | pp8192 | 274.51 ± 0.29 |
|
||||
|
||||
Same error:
|
||||
```
|
||||
ggml_new_object: not enough space in the context's memory pool (needed 26231472, available 26231136)
|
||||
Segmentation fault
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
👤 **saood06** commented the **2025-03-17** at **23:48:58**:<br>
|
||||
|
||||
> I don't have apples for apples comparison as I went down a rabbit hole after realising I could turn off pipeline parallel and it would also give me more VRAM back (thanks @ubergarm!). But it is massively improved.
|
||||
|
||||
Even without the direct comparison, I'm curious what your at now. Also you probably have fixed it by now but CUDA15 was very unused in here:
|
||||
|
||||
> First model load:
|
||||
>[...]
|
||||
>CUDA15 buffer size = 1424.07 MiB
|
||||
|
||||
---
|
||||
|
||||
👤 **davidsyoung** commented the **2025-03-18** at **00:00:30**:<br>
|
||||
|
||||
Damn, I don’t have it right on me as I closed the laptop (night time here). I do have some data in notes from very early run.
|
||||
|
||||
I was able to get to 24k context, with `-ub 2048`. I believe I could get to 32k, but I was getting some errors when playing with `-amb` lower than 32.
|
||||
|
||||
Here are some very initial runs (this is without disabling pipeline parallelism). This is already quite improved from what I can remember.
|
||||
|
||||
Also, for gpu 16, unfortunately I can’t really use it. I can’t split the layers any bit more evenly (at least with what I’ve tried - it’s a bit of a limitation unfortunately without being able to split by row).
|
||||
|
||||
# Compute Buffer Configuration Comparison
|
||||
|
||||
| Parameter/Variable | Run 1 (`-c 8192 -amb 512`) | Run 2 (`-c 16843 -amb 256`) | Notes/Observations |
|
||||
|-----------------------------|----------------------------|-----------------------------|-----------------------------------------------------------------------------------|
|
||||
| **Context Size (`-c`)** | 8,192 | 16,843 | Context doubled (+106%), directly impacts KV cache size. |
|
||||
| **Attention Mask Buffer (`-amb`)** | 512 | 256 | Reduced by 50%, but total compute buffer still increased. |
|
||||
| **Total Compute Buffer** | 31,178.17 MiB | 38,430 MiB | +23% total memory usage despite smaller `-amb`, driven by larger context. |
|
||||
| **KV Self Size** | 549.00 MiB | 1,098.00 MiB | Doubled due to larger context (KV cache scales with sequence length). |
|
||||
| **CUDA_Host Compute Buffer** | 156.05 MiB | 284.05 MiB | +82% increase, likely due to larger context requiring more host-device transfers. |
|
||||
| **Pipeline Copies (`n_copies`)** | 4 | 4 | Pipeline parallelism unchanged. |
|
||||
|
||||
---
|
||||
|
||||
### Example Device Buffer Changes (MiB):
|
||||
| Device | Run 1 | Run 2 | Change |
|
||||
|----------|----------|----------|----------|
|
||||
| CUDA0 | 1,974.01 | 2,342.01 | +19% |
|
||||
| CUDA8 | 2,196.01 | 2,516.01 | +15% |
|
||||
| CUDA15 | 1,936.03 | 2,256.03 | +16% |
|
||||
|
||||
---
|
||||
|
||||
### Key Findings:
|
||||
1. **Context Size Dominates Memory**: Doubling the context size led to a 23% increase in total compute buffer usage despite halving `-amb`.
|
||||
2. **KV Cache Impact**: The KV self size doubled exactly with context length, confirming linear scaling.
|
||||
|
||||
---
|
||||
|
||||
👤 **ikawrakow** commented the **2025-03-18** at **06:36:37**:<br>
|
||||
|
||||
@ubergarm @davidsyoung
|
||||
|
||||
Thank you for testing! It looks like a winner, so merging it.
|
||||
Reference in New Issue
Block a user