qwen3.5 configs (#3554) [skip ci]

* qwen3.5  configs

* update shared experts readme

examples/qwen3.5/122b-a10b-moe-qlora-fsdp.yaml

@@ -31,10 +31,11 @@ lora_target_modules:
   - k_proj
   - v_proj
   - o_proj
-# Regex matching to target shared experts too
-# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
 
-# Target experts
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
 # lora_target_parameters:
 #   - mlp.experts.gate_up_proj
 #   - mlp.experts.down_proj

examples/qwen3.5/122b-a10b-moe-qlora.yaml

@@ -31,11 +31,11 @@ lora_target_modules:
   - k_proj
   - v_proj
   - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
 
-# Regex matching to target shared experts too
-# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
-
-# Target experts
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
 # lora_target_parameters:
 #   - mlp.experts.gate_up_proj
 #   - mlp.experts.down_proj

examples/qwen3.5/35b-a3b-moe-qlora-fsdp.yaml

@@ -31,11 +31,11 @@ lora_target_modules:
   - k_proj
   - v_proj
   - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
 
-# Regex matching to target shared experts too
-# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
-
-# Target experts
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
 # lora_target_parameters:
 #   - mlp.experts.gate_up_proj
 #   - mlp.experts.down_proj

examples/qwen3.5/35b-a3b-moe-qlora.yaml

@@ -42,14 +42,14 @@ lora_target_modules:
   - k_proj
   - v_proj
   - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
 
-# Regex matching to target shared experts too
-# lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
-
-# Target experts
-lora_target_parameters:
-  - mlp.experts.gate_up_proj
-  - mlp.experts.down_proj
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
 
 lora_qkv_kernel: true
 lora_o_kernel: true

examples/qwen3.5/README.md

@@ -59,12 +59,21 @@ lora_target_parameters:
 
 ### Shared Experts (MoE)
 
-Routed experts and shared experts both have `gate_up_proj`/`down_proj`, so a plain module name in `lora_target_modules` would match both. Use a regex to target only attention and shared expert projections, while `lora_target_parameters` above handles routed experts separately:
+Shared experts use `nn.Linear` (unlike routed experts which are 3D `nn.Parameter` tensors), so they can be targeted via `lora_target_modules`. To also train shared expert projections alongside attention, uncomment `gate_up_proj` and `down_proj` in `lora_target_modules`:
 
 ```yaml
-lora_target_modules: 'model\.(language_model\.)?layers\.[\d]+\.(mlp|self_attn)\.(shared_expert\.)?(up|down|gate|gate_up|q|k|v|o)_proj'
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
 ```
 
+Use `lora_target_parameters` (see [Routed Experts](#routed-experts-moe) above) to target routed experts separately.
+
 ### TIPS
 
 - For inference hyp, please see the respective model card details.