tag:github.com,2008:https://github.com/TNTran92/bitsandbytes/releasesRelease notes from bitsandbytes2024-03-15T19:41:33Ztag:github.com,2008:Repository/766030121/rocm-bitsandbyte-0.432024-03-22T20:16:15Zrocm-bitsandbytes-0.43<p>This is a ROCM port of bitsandbytes 0.43</p>
<p>Supported architecture</p>
<p><code>gfx1100</code> family (7900 XTX, 7900 XT, 7900 GRE, etc).</p>
<p>For <code>gfx1101</code> (7700XT, 7800XT) and <code>gfx1102</code> (7600, 7600XT) as well as other not officially supported card, user can try setting the following flag in pyTorch</p>
<p>HSA_OVERRIDE_GFX_VERSION=11.0.0.<br>
<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="2139648547" data-permission-text="Title is private" data-url="https://github.com/ROCm/ROCm/issues/2901" data-hovercard-type="issue" data-hovercard-url="/ROCm/ROCm/issues/2901/hovercard?comment_id=1950136950&comment_type=issue_comment" href="https://github.com/ROCm/ROCm/issues/2901#issuecomment-1950136950">ROCm/ROCm#2901 (comment)</a></p>
<p>The following is shown as supported by the compiler but not tested</p>
<p>-<code>gfx1030</code> (6800, 6800XT, 6900XT, 6950XT)</p>
<p>-<code>gfx942</code> (MI300)</p>
<p>-<code>gfx908</code> (MI200)</p>
<p>-<code>gfx906</code></p>
<p>-<code>gfx900</code></p>
<p>-<code>gfx90a</code></p>TNTran92tag:github.com,2008:Repository/766030121/bitsandbyte-0.432024-03-15T19:41:33Zbitsandbyte-0.43: [docs] refine optimizers, integrations, etc (#1125)<ul>
<li>
<p>optim, integration</p>
</li>
<li>
<p>toctree</p>
</li>
<li>
<p>feedback</p>
</li>
</ul>stevhliu