File size: 4,384 Bytes
3e5aa34
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
# Performance Summary: Qwen3-30B-A3B REAP and FP8-DYNAMIC RTN Ablations Study

## Overview

This report compares 4 compressed/quantized variants of Qwen3-30B-A3B-Instruct-2507 against the unquantized baseline across 5 evaluation benchmarks.

- **FP8-DYNAMIC**: Quantized with FP8-DYNAMIC RTN only
- **REAP-25pct**: Compressed with REAP (25% sparsity) only
- **FP8-DYNAMIC + REAP-25pct**: Quantized with FP8-DYNAMIC RTN first, then compressed with REAP (25% sparsity)
- **REAP-25pct + FP8-DYNAMIC**: Compressed with REAP (25% sparsity) first, then quantized with FP8-DYNAMIC RTN

**Important Note:** The baseline model was served with tp=2 across 2 GPUs, so its server statistics may appear different from the single-GPU variants. The baseline model size (57 GB) is accurate, but its KV cache size and token capacity may be inflated relative to a single-GPU deployment.

## Model Configurations

| Configuration | Model Size | Size Reduction | Session ID |
|---------------|------------|----------------|------------|
| **Baseline (Unquantized)** | 57.00 GB | — | 07102026028 |
| **FP8-DYNAMIC** | 29.21 GB | 48.8% | 07102026-f817 |
| **REAP-25pct** | 43.43 GB | 23.8% | 07102026-aedc |
| **FP8-DYNAMIC + REAP-25pct** | 22.43 GB | 60.6% | 07102026-da72 |
| **REAP-25pct + FP8-DYNAMIC** | 22.43 GB | 60.6% | 07102026-05ea |

## Accuracy Scores

| Task | Baseline | FP8-DYNAMIC | REAP-25pct | FP8-DYNAMIC + REAP-25pct | REAP-25pct + FP8-DYNAMIC |
|------|----------|-------------|------------|---------------------------|---------------------------|
| **gsm8k_platinum_cot_llama** | 0.9680 | 0.9763 | 0.9586 | 0.9573 | 0.9570 |
| **ifeval** | 0.8945 | 0.8921 | 0.8705 | 0.8733 | 0.8713 |
| **math_500\|0** | 0.8887 | 0.8913 | 0.8893 | 0.8740 | 0.8813 |
| **mmlu_cot_llama** | 0.8422 | 0.8397 | 0.7930 | 0.7910 | 0.7921 |
| **mmlu_pro_chat** | 0.7881 | 0.7848 | 0.7264 | 0.7237 | 0.7253 |
| **Average** | **0.8763** | **0.8768** | **0.8476** | **0.8439** | **0.8454** |

## Accuracy Recovery vs. Baseline

| Task | FP8-DYNAMIC | REAP-25pct | FP8-DYNAMIC + REAP-25pct | REAP-25pct + FP8-DYNAMIC |
|------|-------------|------------|---------------------------|---------------------------|
| **gsm8k_platinum_cot_llama** | 100.86% (+0.0083) | 99.03% (-0.0094) | 98.89% (-0.0107) | 98.86% (-0.0110) |
| **ifeval** | 99.73% (-0.0024) | 97.32% (-0.0240) | 97.63% (-0.0212) | 97.41% (-0.0232) |
| **math_500\|0** | 100.29% (+0.0026) | 100.07% (+0.0006) | 98.35% (-0.0147) | 99.17% (-0.0074) |
| **mmlu_cot_llama** | 99.70% (-0.0025) | 94.16% (-0.0492) | 93.92% (-0.0512) | 94.05% (-0.0501) |
| **mmlu_pro_chat** | 99.58% (-0.0033) | 92.17% (-0.0617) | 91.83% (-0.0644) | 92.03% (-0.0628) |
| **Average** | **100.06% (+0.0005)** | **96.72% (-0.0287)** | **96.30% (-0.0324)** | **96.47% (-0.0309)** |

## Variant Ranking (by Average Recovery)

| Rank | Configuration | Avg Recovery | Avg Δ |
|------|---------------|--------------|-------|
| 1 | **FP8-DYNAMIC** | 100.06% | +0.0005 |
| 2 | **REAP-25pct** | 96.72% | -0.0287 |
| 3 | **REAP-25pct + FP8-DYNAMIC** | 96.47% | -0.0309 |
| 4 | **FP8-DYNAMIC + REAP-25pct** | 96.30% | -0.0324 |

## Resource Utilization

| Metric | Baseline | FP8-DYNAMIC | REAP-25pct | FP8-DYNAMIC + REAP-25pct | REAP-25pct + FP8-DYNAMIC |
|--------|----------|-------------|------------|---------------------------|---------------------------|
| KV Cache Size (GB) | 40.29* | 39.90 | 25.86 | 46.57 | 46.57 |
| KV Cache Tokens | 880,192* | 435,776 | 282,448 | 508,672 | 508,672 |
| Recommended Concurrency | 27.51* | 13.62 | 8.83 | 15.90 | 15.90 |

\* Baseline was served with tp=2 across 2 GPUs; KV cache size and token capacity are inflated relative to single-GPU variants and should not be directly compared.

## Concurrency Issues

| Configuration | Task | Issues |
|---------------|------|--------|
| REAP-25pct | mmlu_pro_chat | max KV cache: 91.2% |

---

**Evaluation Date:** 2026-07-10
**vLLM Version:** 0.24.0
**Max Model Length:** 32,000 tokens
**Source PRs:** [#37](https://github.com/Ryfernandes/llm-evaluation-pipeline-results/pull/37), [#38](https://github.com/Ryfernandes/llm-evaluation-pipeline-results/pull/38), [#39](https://github.com/Ryfernandes/llm-evaluation-pipeline-results/pull/39), [#40](https://github.com/Ryfernandes/llm-evaluation-pipeline-results/pull/40), [#41](https://github.com/Ryfernandes/llm-evaluation-pipeline-results/pull/41)