Results of the first run
Browse files
README.md
CHANGED
|
@@ -71,9 +71,9 @@ before choosing it:
|
|
| 71 |
* **Objective — the engine's drafting loop.** `emcdraft.head.unroll` runs all three draft steps
|
| 72 |
for every position of a sequence in parallel, with exactly the key/value visibility the
|
| 73 |
engine's draft cache has (verified prefix, then the round's own entries). Each step is
|
| 74 |
-
trained on the token the target really emitted
|
| 75 |
-
|
| 76 |
-
because the engine only reaches them in that case.
|
| 77 |
* **FP8 with QAT.** An exact, function-preserving re-scaling (`equalize_`) first moves the big
|
| 78 |
matrices into the range where E4M3 has relative precision; they are then trained through
|
| 79 |
E4M3 rounding with a straight-through estimator, so the compact file is the model that was
|
|
@@ -87,8 +87,29 @@ before choosing it:
|
|
| 87 |
|
| 88 |
## Results
|
| 89 |
|
| 90 |
-
|
| 91 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
|
| 93 |
## Layout
|
| 94 |
|
|
@@ -113,7 +134,7 @@ tests/test_head.py CPU self-test (FP8 codec, unroll vs. step-by-step emulatio
|
|
| 113 |
|
| 114 |
## Reproduce
|
| 115 |
|
| 116 |
-
Linux, one GPU with at least 80 GB, about
|
| 117 |
that machine:
|
| 118 |
|
| 119 |
```python
|
|
@@ -139,9 +160,16 @@ This starts `scripts/supervise.py`, which runs `scripts/pipeline.sh` stage by st
|
|
| 139 |
| eval_main, eval_baseline, eval_alt | full official `eval` (160 dev prompts) of the draft, the baseline and the second variant |
|
| 140 |
|
| 141 |
Every stage resumes after an interruption, and logs, kit results, generated data and the
|
| 142 |
-
draft folders are pushed to the dataset repo as they are produced.
|
| 143 |
-
|
| 144 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
|
| 146 |
## Submission format and rules
|
| 147 |
|
|
|
|
| 71 |
* **Objective — the engine's drafting loop.** `emcdraft.head.unroll` runs all three draft steps
|
| 72 |
for every position of a sequence in parallel, with exactly the key/value visibility the
|
| 73 |
engine's draft cache has (verified prefix, then the round's own entries). Each step is
|
| 74 |
+
trained on the token the target really emitted and also distils the target's full
|
| 75 |
+
distribution for that token; later steps are weighted by whether the earlier ones were
|
| 76 |
+
right, because the engine only reaches them in that case.
|
| 77 |
* **FP8 with QAT.** An exact, function-preserving re-scaling (`equalize_`) first moves the big
|
| 78 |
matrices into the range where E4M3 has relative precision; they are then trained through
|
| 79 |
E4M3 rounding with a straight-through estimator, so the compact file is the model that was
|
|
|
|
| 87 |
|
| 88 |
## Results
|
| 89 |
|
| 90 |
+
Official kit `eval` (vLLM 0.31.0, all 160 dev prompts, K = 3), every row measured in the same
|
| 91 |
+
run on the same **NVIDIA RTX PRO 6000 Blackwell**:
|
| 92 |
+
|
| 93 |
+
| draft | τ | chat / code / math / long-form | TPOT | memory | score, same GPU | score, published baseline |
|
| 94 |
+
|---|---|---|---|---|---|---|
|
| 95 |
+
| baseline (the target's built-in MTP head) | 2.806 | 3.06 / 3.74 / 3.64 / 2.37 | 15.39 ms | 0.917 GB | 100.0 | 96.2 |
|
| 96 |
+
| this draft, bf16 | 2.885 | 3.07 / 3.75 / 3.70 / 2.48 | 15.31 ms | 0.917 GB | 102.3 | 98.5 |
|
| 97 |
+
| **this draft, FP8 (submitted variant)** | **2.868** | 3.08 / 3.78 / 3.68 / 2.45 | 15.27 ms | **0.607 GB** | **114.4** | **110.7** |
|
| 98 |
+
|
| 99 |
+
* *Score, same GPU* uses the baseline as measured above; *score, published baseline* is the
|
| 100 |
+
kit's own number, whose speed term compares against the organisers' TPOT from different
|
| 101 |
+
hardware (that is why the baseline itself gets 96.2 there).
|
| 102 |
+
* The FP8 variant's same-GPU score splits into acceptance 1.034 (×0.5), memory 1.5 (×0.25,
|
| 103 |
+
the cap) and speed 1.008 (×0.25). Most of the gain is the compact storage; fine-tuning added
|
| 104 |
+
+0.06 to τ, mostly in long-form writing (+0.08) and math (+0.04).
|
| 105 |
+
* The offline simulator predicted the dev acceptance within 0.015 of the engine (2.858 vs
|
| 106 |
+
2.868 for the FP8 draft, 2.792 vs 2.806 for the baseline).
|
| 107 |
+
* Run r1 trained on 2.02 M generated tokens (4,820 answers) for 584 steps (13 min); its
|
| 108 |
+
validation τ was still rising when the schedule ended. Total GPU time for the whole pipeline,
|
| 109 |
+
including the three full evaluations: about 2.6 hours.
|
| 110 |
+
|
| 111 |
+
Raw kit outputs, per-stage logs, machine statistics and the training summary are in the
|
| 112 |
+
dataset repo under `runs/r1/`; the exported draft folders under `submissions/r1/`.
|
| 113 |
|
| 114 |
## Layout
|
| 115 |
|
|
|
|
| 134 |
|
| 135 |
## Reproduce
|
| 136 |
|
| 137 |
+
Linux, one GPU with at least 80 GB, about 120 GB of disk. In a Python shell or notebook on
|
| 138 |
that machine:
|
| 139 |
|
| 140 |
```python
|
|
|
|
| 160 |
| eval_main, eval_baseline, eval_alt | full official `eval` (160 dev prompts) of the draft, the baseline and the second variant |
|
| 161 |
|
| 162 |
Every stage resumes after an interruption, and logs, kit results, generated data and the
|
| 163 |
+
draft folders are pushed to the dataset repo as they are produced. Settings in
|
| 164 |
+
`control/run.env` extend a finished run instead of repeating it: `GEN_APPEND=1` and
|
| 165 |
+
`EXTRACT_APPEND=1` keep earlier answers and features, `TRAIN_INIT` starts from an earlier
|
| 166 |
+
head, `RUN_NAME` names the new training run.
|
| 167 |
+
|
| 168 |
+
The engine is the pinned version installed from PyPI rather than the container image; the
|
| 169 |
+
results JSON records this. The pipeline sets `VLLM_USE_FLASHINFER_SAMPLER=0`, so the engine
|
| 170 |
+
uses its built-in sampler instead of compiling FlashInfer's at start-up; decoding is greedy,
|
| 171 |
+
so outputs and draft behaviour are unaffected. `python tests/test_head.py` runs the CPU
|
| 172 |
+
self-test (no GPU, no downloads).
|
| 173 |
|
| 174 |
## Submission format and rules
|
| 175 |
|