精度评分卡

Predictions and speed estimates compared against verified hardware benchmarks and ground-truth configurations.

校准报告

测试套件	测试用例数	带内通过率	中位数误差	GMFE 因子误差	系统偏差	最差情况
显存占用	56	100.0%	3.5%	1.07x	微偏低	D5 (0.65x)
解码速度	47	100.0%	16.5%	1.25x	偏低	T9 (0.30x)
微调速度	18	100.0%	22.0%	1.33x	偏低	T2 (0.46x)

内存预测具有高度确定性且非常准确，而速度估算对运行和硬件配置更为敏感。未命中主要是由于源基准中未指定参数数据所致。

带内通过率

100.0%

预测符合目标误差范围的比例

100%

中位数误差

3.5%

典型的绝对预测误差

GMFE 因子误差

1.07x

几何平均误差因子

系统偏差

微偏低

整体校准偏差方向

ID	场景与配置	测量值	预测值	比例	状态
A1	Llama-3-8B (8.03B) fp16 weights = safetensors size	16.1 GB	16.1 GB	1.00x	通过
A2	Qwen2.5-32B (32.8B) Q4_K_M weights = GGUF file size	19.9 GB	19.5 GB	0.98x	通过
A3	Qwen2.5-32B (32.8B) Q6_K weights = GGUF file size	26.9 GB	27.2 GB	1.01x	通过
A4	Qwen2.5-32B (32.8B) Q8_0 weights = GGUF file size	34.8 GB	33.3 GB	0.96x	通过
A5	Qwen2.5-72B (72.7B) Q4_K_M weights = GGUF file size	47.4 GB	45.3 GB	0.96x	通过
A6	Mixtral 8x7B MoE (46.7B total / 12.9B active) Q4_K_M weights, counted once	26.4 GB	27.6 GB	1.04x	通过
A7	Mistral-7B-Instruct-v0.3 (7.25B) bf16 weights = safetensors size	14.5 GB	14.5 GB	1.00x	通过
A8	Llama-3.2-3B (3.21B) bf16 weights = safetensors size (small model, large vocab)	6.4 GB	6.4 GB	1.00x	通过
A9	Qwen2.5-7B (7.62B) Q4_K_M weights = GGUF file size (large-vocab model at 4-bit)	4.7 GB	4.8 GB	1.02x	通过
A10	Llama-2-7B (6.74B) fp16 weights = safetensors size	13.5 GB	13.5 GB	1.00x	通过
B1	Llama-3-8B KV cache, 8k ctx, fp16 (32L, 8 KV heads, head 128)	1.1 GB	1.1 GB	1.00x	通过
B2	Qwen2.5-32B KV cache, 16k ctx, fp16 (64L, 8 KV heads, head 128)	4.3 GB	4.3 GB	1.00x	通过
B3	Llama-3-70B KV cache, 4k ctx, fp16 (80L, 8 KV heads, head 128)	1.3 GB	1.3 GB	1.00x	通过
B4	DeepSeek-V3 MLA KV cache, 32k ctx, bf16 (61L, 576 cached dims/token/layer)	2.3 GB	2.3 GB	1.00x	通过
B5	Mistral-7B-v0.3 KV cache, 32k ctx, fp16 (32L, 8 KV heads, head 128)	4.3 GB	4.3 GB	1.00x	通过
B6	Qwen2.5-72B KV cache, 8k ctx, fp16 (80L, 8 KV heads, head 128)	2.7 GB	2.7 GB	1.00x	通过
B7	Llama-2-7B MHA KV cache, 4k ctx, fp16 (32L, 32 KV heads, head 128)	2.1 GB	2.1 GB	1.00x	通过
B8	Qwen2.5-7B KV cache, 32k ctx, fp16 (28L, 4 KV heads, head 128)	1.9 GB	1.9 GB	1.00x	通过
B9	Llama-3.1-8B KV cache, 128k ctx, fp16 (32L, 8 KV heads, head 128) - long-context stress	17.2 GB	17.2 GB	1.00x	通过
D1	65B QLoRA (nf4, paged optimizer), 1x RTX A6000 (48GB)	44.0 GB	42.4 GB	0.96x	通过
D2	Llama-3.1-8B QLoRA rank 32, batch 1, 2k ctx (~9GB HF+FA2, <8GB Unsloth)	8.5 GB	8.8 GB	1.03x	通过
D3	Llama-3-70B QLoRA (~48GB, fits 1x A100 80GB)	48.0 GB	46.2 GB	0.96x	通过
D4	7B full FT, mixed-precision AdamW (~16 bytes/param static = 112GB before activations)	112.0 GB	130.5 GB	1.17x	通过
D5	70B full FT, 8x H100 (80GB), ZeRO-3, gradient checkpointing, CPU offloading	70.0 GB	45.3 GB	0.65x	通过
D7	Llama-2-7B 16-bit LoRA r=8 (q+v only), micro-batch 1 (measured 21.33GB on A100)	21.3 GB	22.1 GB	1.04x	通过
E1	End-to-end: Llama-3-8B fp16 inference, 2k ctx (measured ~17.5GB loaded)	17.5 GB	17.5 GB	1.00x	通过
E2	End-to-end: Qwen2.5-32B Q4_K_M inference, 16k ctx (measured ~24GB)	24.0 GB	24.8 GB	1.03x	通过
A11	Llama-3.1-405B FP8 weights = 1.0 byte per parameter base weight	405.0 GB	406.2 GB	1.00x	通过
A12	Llama-3.1-405B FP16 weights = 2.0 bytes per parameter base weight	810.0 GB	810.0 GB	1.00x	通过
B10	Llama-3.1-405B KV cache, 131072 ctx, fp8 (126L, 8 KV heads, head 128) - cluster context stress	33.8 GB	33.8 GB	1.00x	通过
B11	DeepSeek-V4-Pro KV cache, 1M ctx, bf16 (61L, 512 shared / 128 indexer dims)	10.3 GB	10.3 GB	1.00x	通过
B12	DeepSeek-V4-Pro KV cache, 1M ctx, fp8/fp4 (61L, 512 shared / 128 indexer dims)	4.7 GB	4.7 GB	1.00x	通过
D8	Llama-3-70B full FT, 32x H100 80GB, ZeRO-3 (no offload)	54.7 GB	43.9 GB	0.80x	通过
E3	End-to-end: Llama-3.1-405B FP8 inference, 8x H100 (80GB), 2k ctx (fits in ~486.8GB VRAM)	486.8 GB	445.6 GB	0.92x	通过
E4	End-to-end: Llama-3.1-405B FP16 inference, 8x H200 (141GB), 2k ctx (fits in ~946.21GB VRAM)	946.2 GB	856.9 GB	0.91x	通过
D9	Llama-2-7B QLoRA NF4 r=8 (q+v), micro-batch 1 (measured 14.18GB on A100)	14.2 GB	13.6 GB	0.96x	通过
D10	Llama-2-7B full FT bf16 mixed, sharded on 2x A100 (measured 36.66GB/GPU)	36.7 GB	32.5 GB	0.89x	通过
D11	Mistral-7B QLoRA r=16, batch 4, 2k ctx, A100-40 (HF+FA2 19.4GB, Unsloth 10.3GB)	12.5 GB	13.3 GB	1.06x	通过
D12	Gemma-7B QLoRA, batch 1, 8k ctx, A100-80 (Unsloth 21.9GB, HF+FA2 47.8GB) - long-seq activation scaling	21.9 GB	17.3 GB	0.79x	通过
D13	Llama-3.1-8B LoRA r=8, batch 2, packed 2k, RTX 4090 (torchtune 16.2 GiB = 17.4GB)	17.4 GB	20.9 GB	1.20x	通过
D14	Llama-3.1-8B QLoRA r=8, batch 2, packed 2k, RTX 4090 (torchtune 7.4 GiB = 7.9GB)	7.9 GB	10.5 GB	1.33x	通过
D15	Llama-3.1-70B LoRA, 8x A100 FSDP full-shard (torchtune 27.6 GiB = 29.6GB/GPU)	29.6 GB	25.4 GB	0.86x	通过
D16	Llama-3.1-405B QLoRA, 8x A100 FSDP (torchtune 44.8GB/GPU)	44.8 GB	49.4 GB	1.10x	通过
D17	Llama-3.1-8B LoRA r=16 (q+v), batch 2, seq 512, NO checkpointing/FA (measured ~24.9GB, RTX PRO 5000 48GB)	24.9 GB	23.9 GB	0.96x	通过
D18	33B QLoRA (nf4, paged optimizer), 1x RTX 4090 (24GB) - Guanaco-33B	21.0 GB	20.9 GB	0.99x	通过
E5	PK 35: Qwen3.5-9B Q4_K_M, INT8 KV, RTX 5070 Ti (16GB), 8k ctx	8.3 GB	7.0 GB	0.84x	通过
E6	PK 37: Qwen3.5-35B-A3B nvfp6, RTX 5090 (32GB), 1k ctx	31.6 GB	28.3 GB	0.89x	通过
E7	PK 38: Qwen3.5-397B-A17B nvfp4, 8x RTX 6000 Blackwell (96GB), 65k ctx	475.2 GB	351.1 GB	0.74x	通过
E8	PK 44: DeepSeek-R1 671B Q8, FP8 KV, 8x B200 (192GB), 131k ctx	1056.1 GB	883.5 GB	0.84x	通过
E9	PK 58: Qwen3.5-122B-A10B Q8, 8x RTX 5000 Blackwell (48GB), 20k ctx	215.4 GB	239.6 GB	1.11x	通过
A13	Qwen3-Omni-30B-A3B BF16 weights = safetensors size	66.0 GB	69.0 GB	1.05x	通过
A14	Qwen3-Omni-30B-A3B W4A16 weights = AutoRound int4 size	25.0 GB	20.5 GB	0.82x	通过
E10	End-to-end: Kimi K3 MXFP4 on 8x B300 (2.8T total parameters, 16/896 MoE)	2100.0 GB	2073.6 GB	0.99x	通过
B10	Llama-3.1-405B KV cache, 131072 ctx, fp8 (126L, 8 KV heads, head 128) - cluster context stress	85.9 GB	85.9 GB	1.00x	通过
D19	DeepSpeed ZeRO-3 Full Finetuning of a 70B model with batch size 16 on 16x H100 (2 nodes)	76.5 GB	81.0 GB	1.06x	通过
E11	Llama-3.1-405B FP8 Inference on 8x H100 (TP=8) with batch size 64	730.0 GB	730.0 GB	1.00x	通过