Adapter relation, Qwen3.5 size tags, Strix Halo hardware line
Browse files
README.md
CHANGED
|
@@ -4,11 +4,18 @@ library_name: llama.cpp
|
|
| 4 |
pipeline_tag: text-generation
|
| 5 |
base_model:
|
| 6 |
- Qwen/Qwen3.5-9B
|
|
|
|
| 7 |
tags:
|
| 8 |
- speculative-decoding
|
| 9 |
-
-
|
| 10 |
- drafter
|
|
|
|
| 11 |
- code
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
---
|
| 13 |
|
| 14 |
# MemDraft
|
|
@@ -30,7 +37,8 @@ One drafter, built from Qwen3.5-9B's own text, used unchanged on four Qwen3.5 si
|
|
| 30 |
| Qwen3.5-4B | 65 tok/s | 110 tok/s (1.69×) | 199 tok/s (3.04×) |
|
| 31 |
| Qwen3.5-9B | 38 tok/s | 70 tok/s (1.84×) | 113 tok/s (2.96×) |
|
| 32 |
|
| 33 |
-
- **Setup:** `ngram-simple`, then MemDraft with the reranker, drafting 3 tokens. Q4_K_M models
|
|
|
|
| 34 |
- **Prompts:** 12 held-out OSS-Instruct coding prompts (2 rounds) and HumanEval problems 0–11. No table holds answers to them. The speedup is over plain decoding on the same prompts; plain decoding runs within about 1% of the same speed on both sets.
|
| 35 |
- **Output:** the target model checks every drafted token, so the text is the target's own. See [Quality](#quality).
|
| 36 |
- **Limits:** Qwen3.5 only, code only, one GPU. The 27B and 35B-A3B are not measured yet. On chat, an earlier table gave only 1.09×.
|
|
|
|
| 4 |
pipeline_tag: text-generation
|
| 5 |
base_model:
|
| 6 |
- Qwen/Qwen3.5-9B
|
| 7 |
+
base_model_relation: adapter
|
| 8 |
tags:
|
| 9 |
- speculative-decoding
|
| 10 |
+
- draft-model
|
| 11 |
- drafter
|
| 12 |
+
- llama.cpp
|
| 13 |
- code
|
| 14 |
+
- qwen3.5
|
| 15 |
+
- qwen3.5-0.8b
|
| 16 |
+
- qwen3.5-2b
|
| 17 |
+
- qwen3.5-4b
|
| 18 |
+
- qwen3.5-9b
|
| 19 |
---
|
| 20 |
|
| 21 |
# MemDraft
|
|
|
|
| 37 |
| Qwen3.5-4B | 65 tok/s | 110 tok/s (1.69×) | 199 tok/s (3.04×) |
|
| 38 |
| Qwen3.5-9B | 38 tok/s | 70 tok/s (1.84×) | 113 tok/s (2.96×) |
|
| 39 |
|
| 40 |
+
- **Setup:** `ngram-simple`, then MemDraft with the reranker, drafting 3 tokens. Q4_K_M models through llama.cpp's Vulkan backend, greedy decoding, 2026-11-21.
|
| 41 |
+
- **Hardware:** an AMD Strix Halo mini PC (Ryzen AI Max+ 395, Radeon 8060S integrated GPU, 128 GB of memory shared by CPU and GPU) on Linux.
|
| 42 |
- **Prompts:** 12 held-out OSS-Instruct coding prompts (2 rounds) and HumanEval problems 0–11. No table holds answers to them. The speedup is over plain decoding on the same prompts; plain decoding runs within about 1% of the same speed on both sets.
|
| 43 |
- **Output:** the target model checks every drafted token, so the text is the target's own. See [Quality](#quality).
|
| 44 |
- **Limits:** Qwen3.5 only, code only, one GPU. The 27B and 35B-A3B are not measured yet. On chat, an earlier table gave only 1.09×.
|