Deploying to gh-pages from @ 3021b23 🚀

Author: puyuan1996

_modules/lightrft/strategy/utils/broadcast_utils.html

@@ -367,36 +367,44 @@ <h1>Source code for lightrft.strategy.utils.broadcast_utils</h1><div class="high
 <span class="sd">        :param name: Original weight name from training model</span>
 <span class="sd">        :return: Mapped weight name for SGLang</span>
 <span class="sd">        &quot;&quot;&quot;</span>
-        <span class="c1"># Step 0: Handle PEFT/LoRA and other potential wrapping prefixes</span>
+        <span class="c1"># Step 0: Handle PEFT/LoRA wrapping prefixes</span>
         <span class="c1"># PEFT models have weights like base_model.model.&lt;original_name&gt;</span>
-        <span class="c1"># We recursively strip &quot;base_model.model.&quot; or &quot;model.&quot; prefixes until we find</span>
-        <span class="c1"># core components like &quot;visual&quot; or &quot;language_model&quot;</span>
-        <span class="k">while</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;base_model.model.&quot;</span><span class="p">)</span> <span class="ow">or</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;model.&quot;</span><span class="p">):</span>
-            <span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;base_model.model.&quot;</span><span class="p">):</span>
-                <span class="n">name</span> <span class="o">=</span> <span class="n">name</span><span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="s2">&quot;base_model.model.&quot;</span><span class="p">):]</span>
-            <span class="k">elif</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;model.&quot;</span><span class="p">):</span>
-                <span class="c1"># We strip &quot;model.&quot; and let the following steps handle it.</span>
-                <span class="c1"># If &quot;language_model&quot; follows, it will be added back as &quot;model.&quot;</span>
-                <span class="c1"># for SGLang&#39;s expectation.</span>
-                <span class="n">name</span> <span class="o">=</span> <span class="n">name</span><span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="s2">&quot;model.&quot;</span><span class="p">):]</span>
+        <span class="c1"># Strip &quot;base_model.model.&quot; prefix (possibly nested) to get the original name.</span>
+        <span class="k">while</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;base_model.model.&quot;</span><span class="p">):</span>
+            <span class="n">name</span> <span class="o">=</span> <span class="n">name</span><span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="s2">&quot;base_model.model.&quot;</span><span class="p">):]</span>
 
         <span class="c1"># PEFT models also rename original weights to include &quot;.base_layer.&quot;</span>
         <span class="c1"># we need to strip this to match standard weight names</span>
         <span class="n">name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s2">&quot;.base_layer.&quot;</span><span class="p">,</span> <span class="s2">&quot;.&quot;</span><span class="p">)</span>
 
-        <span class="c1"># Step 2: Handle language_model prefix mapping</span>
-        <span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;language_model.&quot;</span><span class="p">):</span>
-            <span class="c1"># Remove &quot;language_model.&quot; prefix</span>
-            <span class="n">name</span> <span class="o">=</span> <span class="n">name</span><span class="p">[</span><span class="mi">15</span><span class="p">:]</span>  <span class="c1"># Remove &quot;language_model.&quot;</span>
-
-            <span class="c1"># For lm_head, keep as is (no &quot;model.&quot; prefix)</span>
-            <span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;lm_head&quot;</span><span class="p">):</span>
-                <span class="k">return</span> <span class="n">name</span>
-
-            <span class="c1"># For other components (embed_tokens, layers, norm), add &quot;model.&quot; prefix</span>
-            <span class="k">return</span> <span class="sa">f</span><span class="s2">&quot;model.</span><span class="si">{</span><span class="n">name</span><span class="si">}</span><span class="s2">&quot;</span>
-
-        <span class="c1"># Step 3: Return as is for other cases (e.g., visual.xxx)</span>
+        <span class="c1"># Step 1: Handle VLM models wrapped by ActorVL</span>
+        <span class="c1"># ActorVL wraps the HF model as self.model, so parameter names get an extra &quot;model.&quot; prefix:</span>
+        <span class="c1">#   Training (ActorVL):  model.visual.xxx,       model.model.layers.xxx, model.lm_head.xxx</span>
+        <span class="c1">#   SGLang expects:      visual.xxx,             model.layers.xxx,       lm_head.xxx</span>
+        <span class="c1"># Also handle the &quot;model.language_model.&quot; pattern (some VLM architectures):</span>
+        <span class="c1">#   Training:            model.language_model.model.layers.xxx</span>
+        <span class="c1">#   SGLang expects:      model.layers.xxx</span>
+        <span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;model.language_model.&quot;</span><span class="p">):</span>
+            <span class="n">inner</span> <span class="o">=</span> <span class="n">name</span><span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="s2">&quot;model.language_model.&quot;</span><span class="p">):]</span>
+            <span class="k">if</span> <span class="n">inner</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;lm_head&quot;</span><span class="p">):</span>
+                <span class="k">return</span> <span class="n">inner</span>
+            <span class="k">return</span> <span class="sa">f</span><span class="s2">&quot;model.</span><span class="si">{</span><span class="n">inner</span><span class="si">}</span><span class="s2">&quot;</span>
+
+        <span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;model.visual.&quot;</span><span class="p">):</span>
+            <span class="k">return</span> <span class="n">name</span><span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="s2">&quot;model.&quot;</span><span class="p">):]</span>
+
+        <span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;model.lm_head&quot;</span><span class="p">):</span>
+            <span class="k">return</span> <span class="n">name</span><span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="s2">&quot;model.&quot;</span><span class="p">):]</span>
+
+        <span class="c1"># Handle VLM&#39;s double &quot;model.model.&quot; prefix (ActorVL.model -&gt; HF model.layers)</span>
+        <span class="c1"># model.model.layers.xxx -&gt; model.layers.xxx</span>
+        <span class="c1"># model.model.embed_tokens.xxx -&gt; model.embed_tokens.xxx</span>
+        <span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">&quot;model.model.&quot;</span><span class="p">):</span>
+            <span class="k">return</span> <span class="n">name</span><span class="p">[</span><span class="nb">len</span><span class="p">(</span><span class="s2">&quot;model.&quot;</span><span class="p">):]</span>
+
+        <span class="c1"># Step 2: For text-only models (e.g., Qwen2.5-0.5B-Instruct), parameter names</span>
+        <span class="c1"># are already in SGLang&#39;s expected format: model.layers.xxx, model.embed_tokens.xxx,</span>
+        <span class="c1"># model.norm.xxx, lm_head.xxx. Return as-is without stripping &quot;model.&quot; prefix.</span>
         <span class="k">return</span> <span class="n">name</span>
 
     <span class="k">def</span><span class="w"> </span><span class="nf">_deepspeed_broadcast</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>