Spaces:
Running
Running
Upload roofline.html with huggingface_hub
Browse files- roofline.html +2 -2
roofline.html
CHANGED
|
@@ -83,8 +83,8 @@ fn main(@builtin(global_invocation_id) gid:vec3<u32>){
|
|
| 83 |
<tr><td>Kernel headroom to roofline</td><td class="n">${gap.toFixed(1)}×</td></tr>
|
| 84 |
</table></div>
|
| 85 |
<div class="verdict ${reachable?'ok':'warn'}">${reachable
|
| 86 |
-
? '✓ >1000 tok/s is within
|
| 87 |
-
: '⚠
|
| 88 |
</div>`;
|
| 89 |
}catch(e){ say("✗ "+(e && (e.message||e)), "no"); }
|
| 90 |
})();
|
|
|
|
| 83 |
<tr><td>Kernel headroom to roofline</td><td class="n">${gap.toFixed(1)}×</td></tr>
|
| 84 |
</table></div>
|
| 85 |
<div class="verdict ${reachable?'ok':'warn'}">${reachable
|
| 86 |
+
? '✓ >1000 tok/s is within your GPU roofline ('+roofTok.toFixed(0)+') — a bandwidth-optimal ternary kernel ('+gap.toFixed(1)+'× headroom) gets there; spec-decode is margin.'
|
| 87 |
+
: '⚠ Your GPU batch-1 roofline is '+roofTok.toFixed(0)+' tok/s. A perfect kernel reaches ~'+roofTok.toFixed(0)+'; >1000 needs speculative decode accepting ≥'+specNeeded.toFixed(1)+' tokens/pass, or a card with more bandwidth.'}
|
| 88 |
</div>`;
|
| 89 |
}catch(e){ say("✗ "+(e && (e.message||e)), "no"); }
|
| 90 |
})();
|