address PR code review

- add hf tiny model coverage for Ring 2.0 and Llama 4 adapters

- broaden bailing adapter detection for ring configs

src/axolotl/integrations/aux_free_router/README.md

@@ -37,4 +37,14 @@ Compatibility
 
 Notes
 - If you also enable Liger’s aux-loss paths, the plugin neutralizes aux loss when aux-free is on.
-- Telemetry: future updates will log per-expert loads and bias magnitudes.
+- Telemetry: logs per-layer min/mean/max token loads, `|bias| max`, and bias sign flip fraction using the Trainer’s `logging_steps` cadence.
+- Sample packing: packed batches are compatible with aux-free routing. Because load counts are accumulated on-device per expert before reduction, packing tends to smooth token histograms and reduce bias oscillation. Keep `pad_to_sequence_len: true` when packing to preserve the target token budget per expert.
+
+Telemetry metrics
+- `moe_afb/l{idx}_load_min|mean|max`: token frequency per expert after reduction (0–1 range, sums to 1).
+- `moe_afb/l{idx}_bias_abs_max`: absolute maximum of the learned bias for the layer.
+- `moe_afb/l{idx}_bias_sign_flip_frac`: fraction of experts whose bias sign changed since the previous step (simple oscillation indicator).
+
+Usage tips
+- Increase `logging_steps` if router telemetry becomes noisy for large jobs—the plugin follows the Trainer’s logging cadence.
+- Compare aux-free vs. aux-loss load metrics by plotting the `load_*` series; aux-free typically tightens min/max spread without the auxiliary loss term.

src/axolotl/integrations/aux_free_router/__init__.py

@@ -1,2 +1,9 @@
 """Aux-loss-free (AFB) MoE router integration package."""
 
+from .args import AuxFreeRouterArgs
+from .plugin import AuxFreeMoEPlugin
+
+__all__ = [
+    "AuxFreeMoEPlugin",
+    "AuxFreeRouterArgs",
+]

src/axolotl/integrations/aux_free_router/adapters.py

@@ -1,11 +1,19 @@
+"""Architecture-specific adapters for aux-loss-free MoE routing.
+
+Each adapter discovers MoE layers for a model family and patches only the
+router/gate to inject per-expert bias into expert selection while keeping
+mixture weights from unbiased logits.  Expert dispatch is left untouched so
+the patching composes with any expert backend (eager, ScatterMoE, SonicMoE).
+"""
+
 from __future__ import annotations
 
 from dataclasses import dataclass
 from typing import Iterable, Optional
 
 import torch
-from torch import nn
 import torch.nn.functional as F
+from torch import nn
 
 from axolotl.utils.logging import get_logger
 
@@ -23,9 +31,10 @@ class LayerHandle:
 
 
 class BaseMoEAdapter:
-    """Base adapter that discovers MoE layers and wraps their forward.
+    """Base adapter that discovers MoE layers and patches their routing.
 
-    Concrete adapters should implement discovery and per-layer attribute extraction.
+    Concrete adapters implement discovery, attribute extraction, and
+    architecture-specific router patching.
     """
 
     family: str = "generic"
@@ -33,164 +42,215 @@ class BaseMoEAdapter:
     def matches(self, model: nn.Module) -> bool:  # pragma: no cover - thin shim
         return False
 
-    def find_moe_layers(self, model: nn.Module) -> Iterable[nn.Module]:  # pragma: no cover
+    def find_moe_layers(
+        self, model: nn.Module
+    ) -> Iterable[nn.Module]:  # pragma: no cover
         return []
 
-    def get_top_k(self, moe_layer: nn.Module) -> int:  # pragma: no cover
+    def get_top_k(self, moe_layer: nn.Module) -> int:
-        return int(getattr(moe_layer, "num_experts_per_tok", getattr(moe_layer, "top_k", 2)))
+        """Resolve top_k from the MoE layer, checking common attribute paths."""
+        for attr_path in [
+            ("top_k",),
+            ("num_experts_per_tok",),
+            ("gate", "top_k"),
+            ("router", "top_k"),
+        ]:
+            obj: object = moe_layer
+            for attr in attr_path:
+                obj = getattr(obj, attr, None)
+                if obj is None:
+                    break
+            if isinstance(obj, int):
+                return obj
+        return 2
 
-    def get_num_experts(self, moe_layer: nn.Module) -> int:  # pragma: no cover
+    def get_num_experts(self, moe_layer: nn.Module) -> int:
-        return int(getattr(moe_layer, "num_experts"))
+        """Resolve num_experts from the MoE layer, checking common attribute paths."""
+        for attr_path in [
+            ("num_experts",),
+            ("num_local_experts",),
+            ("gate", "num_experts"),
+            ("router", "num_experts"),
+            ("experts", "num_experts"),
+        ]:
+            obj: object = moe_layer
+            for attr in attr_path:
+                obj = getattr(obj, attr, None)
+                if obj is None:
+                    break
+            if isinstance(obj, int):
+                return obj
+        raise AttributeError(f"Cannot determine num_experts for {type(moe_layer)}")
 
     def disable_aux_loss(self, model_or_layer: nn.Module) -> None:
         # Best-effort: zero router aux loss coef if present
         if hasattr(model_or_layer, "router_aux_loss_coef"):
             try:
-                setattr(model_or_layer, "router_aux_loss_coef", 0.0)
+                model_or_layer.router_aux_loss_coef = 0.0
             except Exception:  # pragma: no cover - non-critical
-                pass
+                LOG.debug(
+                    "disable_aux_loss: failed to set router_aux_loss_coef on %s",
+                    type(model_or_layer).__name__,
+                    exc_info=True,
+                )
 
-    def _register_aux_buffers(self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim) -> None:
+    def _register_aux_buffers(
-        device = next(moe_layer.parameters(), torch.tensor(0)).device
+        self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim
+    ) -> None:
+        p = next(moe_layer.parameters(), None)
+        b = next(moe_layer.buffers(), None)
+        device = (
+            p.device
+            if p is not None
+            else (b.device if b is not None else torch.device("cpu"))
+        )
         if not hasattr(moe_layer, "_afb_bias"):
-            moe_layer.register_buffer("_afb_bias", torch.zeros(handle.num_experts, device=device))
+            moe_layer.register_buffer(
+                "_afb_bias", torch.zeros(handle.num_experts, device=device)
+            )
         if not hasattr(moe_layer, "_afb_counts"):
-            moe_layer.register_buffer("_afb_counts", torch.zeros(handle.num_experts, device=device))
+            moe_layer.register_buffer(
+                "_afb_counts", torch.zeros(handle.num_experts, device=device)
+            )
         if not hasattr(moe_layer, "_afb_ema"):
-            moe_layer.register_buffer("_afb_ema", torch.zeros(handle.num_experts, device=device))
+            moe_layer.register_buffer(
+                "_afb_ema", torch.zeros(handle.num_experts, device=device)
+            )
         moe_layer._afb_layer_idx = handle.layer_idx  # type: ignore[attr-defined]
         moe_layer._afb_top_k = handle.top_k  # type: ignore[attr-defined]
         shim.register_layer_buffers(handle.layer_idx, moe_layer)
 
-    def prepare(self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim) -> None:
+    def prepare(
-        """Attach per-layer buffers and mark as aux-free enabled."""
+        self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim
+    ) -> None:
+        """Attach per-layer buffers.  Subclasses override to also patch routing."""
         self._register_aux_buffers(moe_layer, handle, shim)
-        self._patch_forward_with_aux_free(moe_layer)
 
-    def _patch_forward_with_aux_free(self, moe_layer: nn.Module) -> None:
+    def uses_kernel_routing(self, moe_layer: nn.Module) -> bool:
-        """Replace the layer's forward with an aux-free gating version.
+        """Return True when a kernel backend (SonicMoE / ScatterMoE) has
-
+        already replaced the block forward, meaning the routing is handled
-        Assumes the layer exposes attributes:
+        inside the kernel forward and we should NOT patch the router."""
-          - gate: linear router projecting hidden to num_experts
+        cls = type(moe_layer)
-          - num_experts: int
+        # SonicMoE stores the original forward when it patches a class.
-          - experts: iterable of expert modules taking (tokens, H) -> (tokens, H)
+        if hasattr(cls, "_original_forward"):
-        """
+            return True
-        if getattr(moe_layer, "_afb_patched", False):
+        # ScatterMoE replaces via kernels library; check for the marker.
-            return
+        if hasattr(cls, "_kernel_forward"):
-
+            return True
-        if not hasattr(moe_layer, "gate") or not hasattr(moe_layer, "experts"):
+        return False
-            LOG.info("AuxFreeMoE: layer missing gate/experts; skipping forward patch")
-            return
-
-        def afb_forward(self, hidden_states: torch.Tensor):  # type: ignore[no-redef]
-            # hidden_states: (B, T, H)
-            bsz, seqlen, hdim = hidden_states.shape
-            hs = hidden_states.view(-1, hdim)
-            logits = self.gate(hs)
-            # selection uses biased logits; weights from unbiased logits
-            bias = getattr(self, "_afb_bias")
-            top_k = int(getattr(self, "_afb_top_k", 2))
-            biased = logits + bias  # broadcast over tokens
-            topk_vals, topk_idx = torch.topk(biased, k=top_k, dim=-1, sorted=False)
-            chosen_logits = torch.gather(logits, -1, topk_idx)
-            weights = torch.softmax(chosen_logits.float(), dim=-1)
-            weights = weights.to(hs.dtype)
-
-            # accumulate counts for bias update callback
-            flat_idx = topk_idx.reshape(-1)
-            counts = torch.bincount(flat_idx, minlength=int(self.num_experts))
-            getattr(self, "_afb_counts").add_(counts.to(getattr(self, "_afb_counts").dtype))
-
-            # dispatch tokens to experts
-            hs_rep = hs.repeat_interleave(top_k, dim=0)
-            y = torch.empty_like(hs_rep)
-            for eid in range(int(self.num_experts)):
-                mask = flat_idx == eid
-                if mask.any():
-                    y[mask] = self.experts[eid](hs_rep[mask])
-
-            y = (y.view(-1, top_k, hdim) * weights.unsqueeze(-1)).sum(dim=1)
-            out = y.view(bsz, seqlen, hdim)
-            return (out, logits)
-
-        moe_layer.forward = afb_forward.__get__(moe_layer, moe_layer.__class__)  # type: ignore[attr-defined]
-        setattr(moe_layer, "_afb_patched", True)
 
 
 class MixtralAdapter(BaseMoEAdapter):
+    """Patches the TopKRouter for Mixtral / Qwen-MoE style softmax→topk
+    routing so that biased logits drive expert *selection* while unbiased
+    softmax scores drive mixture *weights*.
+
+    Works with transformers v5 where experts are fused 3D tensors and
+    the router is ``MixtralTopKRouter`` (returns a 3-tuple).
+    """
+
     family = "mixtral"
 
     def matches(self, model: nn.Module) -> bool:
-        return getattr(getattr(model, "config", object()), "model_type", "") == "mixtral"
+        return (
-
+            getattr(getattr(model, "config", object()), "model_type", "") == "mixtral"
-    def prepare(self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim) -> None:
+        )
-        self._register_aux_buffers(moe_layer, handle, shim)
-        self._patch_mixtral_forward(moe_layer, shim)
 
     def find_moe_layers(self, model: nn.Module) -> Iterable[nn.Module]:
         for m in model.modules():
             if m.__class__.__name__.endswith("SparseMoeBlock"):
                 yield m
 
-    def _patch_mixtral_forward(self, moe_layer: nn.Module, shim: AuxFreeShim) -> None:
+    def prepare(
-        if getattr(moe_layer, "_afb_patched", False):
+        self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim
-            return
+    ) -> None:
-
+        self._register_aux_buffers(moe_layer, handle, shim)
-        shim_ref = shim
+        if not self.uses_kernel_routing(moe_layer):
-
+            self._patch_router(moe_layer)
-        def afb_forward(self, hidden_states: torch.Tensor):  # type: ignore[no-redef]
+        else:
-            batch_size, sequence_length, hidden_dim = hidden_states.shape
+            LOG.info(
-            if self.training and getattr(self, "jitter_noise", 0) > 0:
+                "AuxFreeMoE: kernel backend detected on %s; "
-                hidden_states = hidden_states * torch.empty_like(hidden_states).uniform_(
+                "skipping router patch (kernel routing handles bias)",
-                    1.0 - self.jitter_noise, 1.0 + self.jitter_noise
+                type(moe_layer).__name__,
-                )
-            flat_states = hidden_states.view(-1, hidden_dim)
-            router_logits = self.gate(flat_states)
-
-            layer_idx = int(getattr(self, "_afb_layer_idx", 0))
-            top_k = int(getattr(self, "_afb_top_k", self.top_k))
-            selected_experts, routing_weights = shim_ref.select_experts(layer_idx, router_logits, top_k)
-            routing_weights = routing_weights.to(flat_states.dtype)
-
-            flat_idx = selected_experts.reshape(-1)
-            counts = torch.bincount(flat_idx, minlength=int(self.num_experts))
-            self._afb_counts.add_(counts.to(self._afb_counts.dtype))
-
-            final_hidden_states = torch.zeros(
-                (batch_size * sequence_length, hidden_dim),
-                dtype=flat_states.dtype,
-                device=flat_states.device,
             )
 
-            expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+    def _patch_router(self, moe_layer: nn.Module) -> None:
-            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+        """Patch the TopKRouter to inject aux-free bias into expert selection."""
-            for expert_idx in expert_hit:
+        gate = getattr(moe_layer, "gate", None)
-                expert_layer = self.experts[expert_idx]
+        if gate is None:
-                idx, top_x = torch.where(expert_mask[expert_idx].squeeze(0))
+            LOG.info("MixtralAdapter: layer missing gate; skipping aux-free patch")
-                current_state = flat_states[None, top_x].reshape(-1, hidden_dim)
+            return
-                current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+        if getattr(gate, "_afb_patched", False):
-                final_hidden_states.index_add_(0, top_x, current_hidden_states.to(flat_states.dtype))
+            return
 
-            final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        # Capture reference to the MoE block for bias / counts access.
-            return final_hidden_states, router_logits
+        block_ref = moe_layer
 
-        moe_layer.forward = afb_forward.__get__(moe_layer, moe_layer.__class__)  # type: ignore[attr-defined]
+        def afb_router_forward(self, hidden_states: torch.Tensor):
-        setattr(moe_layer, "_afb_patched", True)
+            hidden_states = hidden_states.reshape(-1, self.hidden_dim)
+            router_logits = F.linear(hidden_states, self.weight)
+            router_probs = F.softmax(router_logits.float(), dim=-1)
+
+            # Biased selection, unbiased weights
+            bias = block_ref._afb_bias
+            biased = router_probs + bias
+            _, router_indices = torch.topk(biased, self.top_k, dim=-1)
+            router_scores = torch.gather(router_probs, 1, router_indices)
+
+            # Renormalize (Mixtral always normalizes; Qwen checks config)
+            if getattr(self, "norm_topk_prob", True):
+                router_scores = router_scores / router_scores.sum(dim=-1, keepdim=True)
+
+            # Accumulate counts for the bias-update callback
+            flat_idx = router_indices.reshape(-1)
+            counts = torch.bincount(flat_idx, minlength=self.num_experts)
+            block_ref._afb_counts.add_(counts.to(block_ref._afb_counts.dtype))
+
+            return router_probs, router_scores, router_indices
+
+        gate.forward = afb_router_forward.__get__(gate, gate.__class__)  # type: ignore[attr-defined]
+        gate._afb_patched = True
+        moe_layer._afb_patched = True
 
 
 class Qwen3Adapter(MixtralAdapter):
     family = "qwen3_moe"
 
     def matches(self, model: nn.Module) -> bool:
-        return getattr(getattr(model, "config", object()), "model_type", "") in ("qwen3_moe", "qwen2_moe")
+        return getattr(getattr(model, "config", object()), "model_type", "") in (
+            "qwen3_moe",
+            "qwen2_moe",
+        )
+
+
+class Qwen35MoeAdapter(MixtralAdapter):
+    """Adapter for Qwen 3.5 MoE models.
+
+    Same softmax→topk router pattern as Mixtral/Qwen3.  The shared expert
+    is handled by the block forward (untouched by router-level patching).
+    """
+
+    family = "qwen3_5_moe"
+
+    def matches(self, model: nn.Module) -> bool:
+        return getattr(getattr(model, "config", object()), "model_type", "") in (
+            "qwen3_5_moe",
+            "qwen3_5_moe_text",
+        )
 
 
 class BailingAdapter(BaseMoEAdapter):
     family = "bailing_moe"
 
     def matches(self, model: nn.Module) -> bool:
-        model_type = getattr(getattr(model, "config", object()), "model_type", "")
+        cfg = getattr(model, "config", None)
-        return model_type in ("bailing_moe", "bailing_moe_v2")
+        if cfg is None:
+            return False
+        model_type = getattr(cfg, "model_type", "") or ""
+        if model_type in ("bailing_moe", "bailing_moe_v2", "ring_moe", "ring"):
+            return True
+        cfg_name = cfg.__class__.__name__.lower()
+        return "bailingmoev2" in cfg_name or "ring" in cfg_name
 
     def find_moe_layers(self, model: nn.Module) -> Iterable[nn.Module]:
         for m in model.modules():
@@ -199,11 +259,15 @@ class BailingAdapter(BaseMoEAdapter):
 
     def get_num_experts(self, moe_layer: nn.Module) -> int:
         if hasattr(moe_layer, "num_experts"):
-            return int(getattr(moe_layer, "num_experts"))
+            return int(moe_layer.num_experts)
         cfg = getattr(moe_layer, "config", None)
-        return int(getattr(cfg, "num_experts"))
+        if cfg is None:
+            raise AttributeError(f"Cannot determine num_experts for {type(moe_layer)}")
+        return int(cfg.num_experts)
 
-    def prepare(self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim) -> None:
+    def prepare(
+        self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim
+    ) -> None:
         self._register_aux_buffers(moe_layer, handle, shim)
         self._patch_bailing_gate(moe_layer)
 
@@ -219,9 +283,9 @@ class BailingAdapter(BaseMoEAdapter):
             flat = hidden_states.view(-1, hidden_states.shape[-1])
             logits = F.linear(flat.float(), self.weight.float())
             scores_unbiased = torch.sigmoid(logits.float()).to(logits.dtype)
-            bias = getattr(moe_layer, "_afb_bias")
+            bias = moe_layer._afb_bias
             biased_scores = scores_unbiased + bias
-            topk_vals, topk_idx = self.group_limited_topk(biased_scores)
+            _, topk_idx = self.group_limited_topk(biased_scores)
             weights = torch.gather(scores_unbiased, 1, topk_idx)
             if self.top_k > 1:
                 denom = weights.sum(dim=-1, keepdim=True).clamp_min_(1e-20)
@@ -230,12 +294,12 @@ class BailingAdapter(BaseMoEAdapter):
 
             flat_topk = topk_idx.reshape(-1)
             counts = torch.bincount(flat_topk, minlength=bias.numel())
-            getattr(moe_layer, "_afb_counts").add_(counts.to(moe_layer._afb_counts.dtype))
+            moe_layer._afb_counts.add_(counts.to(moe_layer._afb_counts.dtype))
 
             return topk_idx, weights.to(hidden_states.dtype), logits
 
         gate.forward = afb_gate_forward.__get__(gate, gate.__class__)  # type: ignore[attr-defined]
-        setattr(gate, "_afb_patched", True)
+        gate._afb_patched = True
 
 
 class Llama4Adapter(BaseMoEAdapter):
@@ -249,7 +313,9 @@ class Llama4Adapter(BaseMoEAdapter):
             if m.__class__.__name__ == "Llama4TextMoe":
                 yield m
 
-    def prepare(self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim) -> None:
+    def prepare(
+        self, moe_layer: nn.Module, handle: LayerHandle, shim: AuxFreeShim
+    ) -> None:
         self._register_aux_buffers(moe_layer, handle, shim)
         self._patch_llama4_router(moe_layer)
 
@@ -262,9 +328,13 @@ class Llama4Adapter(BaseMoEAdapter):
             return
 
         def afb_router_forward(self, hidden_states: torch.Tensor):
-            flat = hidden_states if hidden_states.dim() == 2 else hidden_states.view(-1, hidden_states.shape[-1])
+            flat = (
+                hidden_states
+                if hidden_states.dim() == 2
+                else hidden_states.view(-1, hidden_states.shape[-1])
+            )
             router_logits = F.linear(flat, self.weight, self.bias)
-            bias = getattr(moe_layer, "_afb_bias")
+            bias = moe_layer._afb_bias
             biased_logits = router_logits + bias
             _, router_indices = torch.topk(biased_logits, self.top_k, dim=1)
             unbiased_top = torch.gather(router_logits, 1, router_indices)
@@ -273,15 +343,17 @@ class Llama4Adapter(BaseMoEAdapter):
             router_scores = torch.sigmoid(router_scores.float()).to(router_scores.dtype)
 
             counts = torch.bincount(router_indices.reshape(-1), minlength=bias.numel())
-            getattr(moe_layer, "_afb_counts").add_(counts.to(moe_layer._afb_counts.dtype))
+            moe_layer._afb_counts.add_(counts.to(moe_layer._afb_counts.dtype))
 
             return router_scores, router_logits
 
         router.forward = afb_router_forward.__get__(router, router.__class__)  # type: ignore[attr-defined]
-        setattr(router, "_afb_patched", True)
+        router._afb_patched = True
 
 
-def discover_and_prepare_layers(model: nn.Module, adapters: list[BaseMoEAdapter], shim: AuxFreeShim) -> list[LayerHandle]:
+def discover_and_prepare_layers(
+    model: nn.Module, adapters: list[BaseMoEAdapter], shim: AuxFreeShim
+) -> list[LayerHandle]:
     """Discover MoE layers using the first matching adapter and attach per-layer buffers.
 
     Returns a list of layer handles for later routing patching and updates.
@@ -305,7 +377,7 @@ def discover_and_prepare_layers(model: nn.Module, adapters: list[BaseMoEAdapter]
         try:
             top_k = adapter.get_top_k(layer)
             nE = adapter.get_num_experts(layer)
-        except Exception:
+        except (AttributeError, TypeError, ValueError):
             continue
 
         handle = LayerHandle(layer=layer, layer_idx=idx, num_experts=nE, top_k=top_k)
@@ -313,5 +385,9 @@ def discover_and_prepare_layers(model: nn.Module, adapters: list[BaseMoEAdapter]
         handles.append(handle)
         idx += 1
 
-    LOG.info(f"AuxFreeMoE: prepared {len(handles)} {adapter.family} layers for aux-free routing")
+    LOG.info(
+        "AuxFreeMoE: prepared %d %s layers for aux-free routing",
+        len(handles),
+        adapter.family,
+    )
     return handles

src/axolotl/integrations/aux_free_router/args.py

@@ -0,0 +1,71 @@
+# Copyright 2024 Axolotl AI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Plugin args for the Aux-Loss-Free MoE router integration.
+"""
+
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+
+class AuxFreeRouterArgs(BaseModel):
+    """
+    Input args for Aux-Loss-Free MoE routing.
+    """
+
+    moe_balance_type: Literal["gshard", "noaux_tc"] | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "MoE load balancing strategy: 'gshard' for auxiliary loss, "
+            "'noaux_tc' for aux-loss-free bias updates affecting top-k selection only. "
+            "Defaults to model's native behavior when unset."
+        },
+    )
+    moe_update_rate: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Per-step bias update rate (gamma). Recommended: 0.005-0.05. "
+            "If unset, plugin default is 0.01."
+        },
+    )
+    moe_update_momentum: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "EMA momentum for expert load smoothing (0-1). "
+            "If unset, plugin default is 0.9."
+        },
+    )
+    moe_bias_cap: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Absolute clamp for expert bias magnitude. "
+            "If unset, plugin default is 2.0."
+        },
+    )
+    moe_afb_warmup_steps: int | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Number of initial steps to delay aux-free bias updates, "
+            "allowing routing to stabilize. If unset, plugin default is 0."
+        },
+    )
+    moe_bias_sync_group: Literal["world", "ep"] | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Reduction group for expert load counts: 'world' (DP) or "
+            "'ep' (expert-parallel group if available). Defaults to 'world' when unset."
+        },
+    )

src/axolotl/integrations/aux_free_router/core.py

@@ -5,6 +5,7 @@ from typing import Optional
 
 import torch
 import torch.distributed as dist
+
 from axolotl.utils.logging import get_logger
 
 LOG = get_logger(__name__)
@@ -22,9 +23,17 @@ class AuxFreeConfig:
 class AuxFreeState:
     """Holds per-layer bias and EMA load buffers."""
 
-    def __init__(self, num_layers: int, num_experts: int, device: torch.device, cfg: AuxFreeConfig):
+    def __init__(
+        self,
+        num_layers: int,
+        num_experts: int,
+        device: torch.device,
+        cfg: AuxFreeConfig,
+    ):
         self.bias = [torch.zeros(num_experts, device=device) for _ in range(num_layers)]
-        self.ema_load = [torch.zeros(num_experts, device=device) for _ in range(num_layers)]
+        self.ema_load = [
+            torch.zeros(num_experts, device=device) for _ in range(num_layers)
+        ]
         self.cfg = cfg
         self.steps = 0
 
@@ -45,17 +54,20 @@ class AuxFreeShim:
             self.state.cfg.sync_group == "ep" and self.ep_group is None
         )
         self._layer_modules: dict[int, torch.nn.Module] = {}
+        self._prev_bias_sign: dict[int, torch.Tensor] = {}
 
     @torch.no_grad()
-    def select_experts(self, layer_idx: int, logits: torch.Tensor, top_k: int) -> tuple[torch.Tensor, torch.Tensor]:
+    def select_experts(
+        self, layer_idx: int, logits: torch.Tensor, top_k: int
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         """Returns (topk_indices, weights) using biased selection and unbiased weights."""
         module = self._layer_modules.get(layer_idx)
         if module is not None and hasattr(module, "_afb_bias"):
-            b = getattr(module, "_afb_bias")
+            b = module._afb_bias
         else:
             b = self.state.bias[layer_idx]
         biased = logits + b  # bias is a buffer
-        topk_scores, topk_idx = torch.topk(biased, k=top_k, dim=-1)
+        _topk_scores, topk_idx = torch.topk(biased, k=top_k, dim=-1)
         chosen_logits = torch.gather(logits, -1, topk_idx)
         weights = torch.softmax(chosen_logits.float(), dim=-1).to(logits.dtype)
         return topk_idx, weights
@@ -63,8 +75,8 @@ class AuxFreeShim:
     def register_layer_buffers(self, layer_idx: int, module: torch.nn.Module) -> None:
         """Bind model buffers so shim updates stay in sync with patched layers."""
         self._layer_modules[layer_idx] = module
-        bias = getattr(module, "_afb_bias")
+        bias = module._afb_bias
-        ema = getattr(module, "_afb_ema")
+        ema = module._afb_ema
         # Keep state views pointing to the same tensors to avoid drift.
         if layer_idx < len(self.state.bias):
             self.state.bias[layer_idx] = bias
@@ -75,6 +87,9 @@ class AuxFreeShim:
         """Call once per optimizer step before per-layer updates."""
         self.state.steps += 1
 
+    def get_prev_bias_sign(self, layer_idx: int) -> Optional[torch.Tensor]:
+        return self._prev_bias_sign.get(layer_idx)
+
     @torch.no_grad()
     def all_reduce_counts(self, counts: torch.Tensor) -> torch.Tensor:
         self._maybe_init_ep_group()
@@ -96,8 +111,8 @@ class AuxFreeShim:
             return
         module = self._layer_modules.get(layer_idx)
         if module is not None and hasattr(module, "_afb_ema"):
-            ema = getattr(module, "_afb_ema")
+            ema = module._afb_ema
-            bias = getattr(module, "_afb_bias")
+            bias = module._afb_bias
         else:
             ema = self.state.ema_load[layer_idx]
             bias = self.state.bias[layer_idx]
@@ -111,6 +126,7 @@ class AuxFreeShim:
         bias.add_(delta)
         if cfg.bias_cap is not None and cfg.bias_cap > 0:
             bias.clamp_(-cfg.bias_cap, cfg.bias_cap)
+        self._prev_bias_sign[layer_idx] = torch.sign(bias.detach())
 
     def _maybe_init_ep_group(self) -> None:
         if not self._ep_group_pending:

src/axolotl/integrations/aux_free_router/plugin.py

@@ -6,7 +6,7 @@ unbiased logits for mixture weights and per-expert biases for top-k selection.
 
 from __future__ import annotations
 
-from typing import Optional
+from typing import Any, Optional
 
 import torch
 import torch.distributed as dist
@@ -21,6 +21,7 @@ from .adapters import (
     Llama4Adapter,
     MixtralAdapter,
     Qwen3Adapter,
+    Qwen35MoeAdapter,
     discover_and_prepare_layers,
 )
 from .core import AuxFreeConfig, AuxFreeShim, AuxFreeState
@@ -29,24 +30,29 @@ LOG = get_logger(__name__)
 
 
 class MoeAuxFreeBiasUpdateCallback(TrainerCallback):
-    """Post-step callback to update aux-free biases from accumulated expert counts.
+    """Post-step callback to update aux-free biases from accumulated expert counts."""
 
-    Note: The current revision expects per-layer counts to be accumulated on each
+    def __init__(
-    MoE layer as a buffer named `_afb_counts` during forward (to be added with
+        self,
-    routing patches in a follow-up).
+        shim: AuxFreeShim,
-    """
+        layer_modules: list[torch.nn.Module],
-
+        trainer: Any,
-    def __init__(self, shim: AuxFreeShim, layer_modules: list[torch.nn.Module]):
+    ):
         self.shim = shim
         self.layer_modules = layer_modules
+        self.trainer = trainer
+        self._prev_bias_sign: dict[int, torch.Tensor] = {}
+        self._telemetry_buffer: dict[int, dict[str, float]] = {}
 
     def on_step_end(self, args, state, control, **kwargs):  # noqa: D401
         # Iterate prepared MoE layers and apply the bias update rule.
         self.shim.begin_step()
         for layer in self.layer_modules:
-            if not hasattr(layer, "_afb_counts") or not hasattr(layer, "_afb_layer_idx"):
+            if not hasattr(layer, "_afb_counts") or not hasattr(
+                layer, "_afb_layer_idx"
+            ):
                 continue
-            counts = getattr(layer, "_afb_counts")
+            counts = layer._afb_counts
             if counts is None:
                 continue
             counts = self.shim.all_reduce_counts(counts)
@@ -54,15 +60,73 @@ class MoeAuxFreeBiasUpdateCallback(TrainerCallback):
             if layer_idx is None:
                 counts.zero_()
                 continue
-            bias = getattr(layer, "_afb_bias")
+            bias = layer._afb_bias
             counts_for_update = counts.to(bias.device)
             tokens_seen = int(counts_for_update.sum().item())
             # local layer-state EMA and bias update
             self.shim.update_bias(layer_idx, counts_for_update, tokens_seen)
+            self._collect_telemetry(layer_idx, counts_for_update, tokens_seen, bias)
             # reset step counts
             counts.zero_()
+
+        if self._should_log(args, state) and self._telemetry_buffer:
+            logs: dict[str, float] = {}
+            for layer_idx, metrics in sorted(self._telemetry_buffer.items()):
+                prefix = f"moe_afb/l{layer_idx}_"
+                for key, value in metrics.items():
+                    logs[f"{prefix}{key}"] = value
+            if logs and hasattr(self.trainer, "log"):
+                self.trainer.log(logs)
+            self._telemetry_buffer.clear()
         return control
 
+    def _collect_telemetry(
+        self,
+        layer_idx: int,
+        counts: torch.Tensor,
+        tokens_seen: int,
+        bias: torch.Tensor,
+    ) -> None:
+        if tokens_seen <= 0:
+            return
+        freq = counts.float() / float(tokens_seen)
+        load_min = freq.min().item()
+        load_mean = freq.mean().item()
+        load_max = freq.max().item()
+        bias_abs_max = bias.abs().max().item()
+
+        prev_sign = self._prev_bias_sign.get(layer_idx)
+        current_sign = torch.sign(bias.detach())
+        if prev_sign is None or prev_sign.shape != current_sign.shape:
+            oscillation = 0.0
+        else:
+            changed = (current_sign != prev_sign) & (
+                (current_sign != 0) | (prev_sign != 0)
+            )
+            if changed.numel() == 0:
+                oscillation = 0.0
+            else:
+                oscillation = changed.float().mean().item()
+        self._prev_bias_sign[layer_idx] = current_sign.clone()
+
+        self._telemetry_buffer[layer_idx] = {
+            "load_min": load_min,
+            "load_mean": load_mean,
+            "load_max": load_max,
+            "bias_abs_max": bias_abs_max,
+            "bias_sign_flip_frac": oscillation,
+        }
+
+    def _should_log(self, args, state) -> bool:
+        interval = getattr(args, "logging_steps", 0)
+        if not interval:
+            return False
+        try:
+            interval = max(1, int(interval))
+        except (TypeError, ValueError):
+            return False
+        return interval > 0 and state.global_step % interval == 0
+
 
 class AuxFreeMoEPlugin(BasePlugin):
     """Plugin that enables aux-loss-free routing when configured."""
@@ -73,18 +137,25 @@ class AuxFreeMoEPlugin(BasePlugin):
         self._shim: Optional[AuxFreeShim] = None
         self._ep_group_cache: dict[tuple[int, ...], dist.ProcessGroup] = {}
 
+    def get_input_args(self):
+        return "axolotl.integrations.aux_free_router.AuxFreeRouterArgs"
+
     def post_model_build(self, cfg, model):
         # Enable only when explicitly requested
         if getattr(cfg, "moe_balance_type", None) != "noaux_tc":
             return
 
         # Be conservative — skip known native aux-free families
-        native_auxfree = getattr(getattr(model, "config", object()), "model_type", "") in (
+        native_auxfree = getattr(
+            getattr(model, "config", object()), "model_type", ""
+        ) in (
             "deepseek_v3",
             "glm4_moe",
         )
         if native_auxfree:
-            LOG.info("AuxFreeMoE: model reports native aux-free routing; skipping patching")
+            LOG.info(
+                "AuxFreeMoE: model reports native aux-free routing; skipping patching"
+            )
             return
 
         # Build aux-free state and shim
@@ -96,13 +167,18 @@ class AuxFreeMoEPlugin(BasePlugin):
         warmup = cfg.moe_afb_warmup_steps if cfg.moe_afb_warmup_steps is not None else 0
         sync_group = cfg.moe_bias_sync_group if cfg.moe_bias_sync_group else "world"
         af_cfg = AuxFreeConfig(
-            rate=rate, momentum=momentum, bias_cap=bias_cap, warmup_steps=warmup, sync_group=sync_group
+            rate=rate,
+            momentum=momentum,
+            bias_cap=bias_cap,
+            warmup_steps=warmup,
+            sync_group=sync_group,
         )
 
         # Discover layers to count the number and experts for state sizing
         adapters: list[BaseMoEAdapter] = [
             MixtralAdapter(),
             Qwen3Adapter(),
+            Qwen35MoeAdapter(),
             BailingAdapter(),
             Llama4Adapter(),
         ]
@@ -110,7 +186,7 @@ class AuxFreeMoEPlugin(BasePlugin):
         # For initial state sizing, we conservatively assume the first discovered layer defines nE
         n_layers = 0
         n_experts = None
-        for m in model.modules():
+        for _m in model.modules():
             n_layers += 1  # upper bound — we will re-use bias slots sparsely
         device = next(model.parameters(), torch.tensor(0)).device
         if n_layers <= 0:
@@ -118,7 +194,9 @@ class AuxFreeMoEPlugin(BasePlugin):
         if n_experts is None:
             # we'll set a minimal placeholder; prepare() will conceptually use module buffers instead
             n_experts = 2
-        state = AuxFreeState(num_layers=n_layers, num_experts=n_experts, device=device, cfg=af_cfg)
+        state = AuxFreeState(
+            num_layers=n_layers, num_experts=n_experts, device=device, cfg=af_cfg
+        )
         ep_size = getattr(cfg, "expert_parallel_size", None)
         ep_group = None
         if sync_group == "ep":
@@ -139,11 +217,15 @@ class AuxFreeMoEPlugin(BasePlugin):
 
     def _resolve_ep_group(self, cfg) -> Optional[dist.ProcessGroup]:
         if not dist.is_available() or not dist.is_initialized():
-            LOG.warning("AuxFreeMoE: EP sync requested but torch.distributed is not initialized; defaulting to world")
+            LOG.warning(
+                "AuxFreeMoE: EP sync requested but torch.distributed is not initialized; defaulting to world"
+            )
             return None
         ep_size = getattr(cfg, "expert_parallel_size", None)
         if not ep_size or ep_size <= 1:
-            LOG.warning("AuxFreeMoE: moe_bias_sync_group='ep' but expert_parallel_size<=1; defaulting to world")
+            LOG.warning(
+                "AuxFreeMoE: moe_bias_sync_group='ep' but expert_parallel_size<=1; defaulting to world"
+            )
             return None
         world = dist.get_world_size()
         if world % ep_size != 0:
@@ -157,11 +239,17 @@ class AuxFreeMoEPlugin(BasePlugin):
             return dist.group.WORLD
 
         rank = dist.get_rank()
-        group_start = (rank // ep_size) * ep_size
+        # All ranks must collectively create all EP subgroups in the same order
-        ranks = tuple(range(group_start, group_start + ep_size))
+        # to avoid deadlocks (dist.new_group is a collective operation).
-        if ranks not in self._ep_group_cache:
+        world_size = world
-            self._ep_group_cache[ranks] = dist.new_group(ranks)
+        my_group = None
-        return self._ep_group_cache[ranks]
+        for group_start in range(0, world_size, ep_size):
+            ranks = tuple(range(group_start, group_start + ep_size))
+            if ranks not in self._ep_group_cache:
+                self._ep_group_cache[ranks] = dist.new_group(ranks)
+            if rank in ranks:
+                my_group = self._ep_group_cache[ranks]
+        return my_group
 
     def add_callbacks_post_trainer(self, cfg, trainer):
         if getattr(cfg, "moe_balance_type", None) != "noaux_tc":
@@ -170,6 +258,10 @@ class AuxFreeMoEPlugin(BasePlugin):
             return []
         # gather concrete layer modules from handles
         layers = [h.layer for h in self._handles]
-        cb = MoeAuxFreeBiasUpdateCallback(self._shim, layers)
+        cb = MoeAuxFreeBiasUpdateCallback(
+            self._shim,
+            layers,
+            trainer,
+        )
         LOG.info("AuxFreeMoE: registering post-step bias update callback")
         return [cb]

src/axolotl/integrations/kernels/libs/scattermoe_lora/layers.py

@@ -240,7 +240,16 @@ def _softmax_topk_route(
 
     top_k = base_gate.top_k
     num_experts = base_gate.num_experts
-    routing_weights, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    # Aux-free bias: biased selection, unbiased weights
+    afb_bias = getattr(moe_block, "_afb_bias", None)
+    if afb_bias is not None:
+        scores_for_choice = routing_weights + afb_bias
+        _, selected_experts = torch.topk(scores_for_choice, top_k, dim=-1)
+        routing_weights = routing_weights.gather(1, selected_experts)
+        _accumulate_afb_counts(moe_block, selected_experts)
+    else:
+        routing_weights, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
 
     if getattr(base_gate, "norm_topk_prob", True):
         routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
@@ -282,6 +291,11 @@ def _sigmoid_topk_route(
     else:
         scores_for_choice = router_probs
 
+    # Aux-free bias: stacks on top of e_score_correction_bias for selection
+    afb_bias = getattr(moe_block, "_afb_bias", None)
+    if afb_bias is not None:
+        scores_for_choice = scores_for_choice + afb_bias
+
     # Group-based selection: pick top groups, mask the rest
     n_group = getattr(moe_block, "n_group", 1)
     if n_group > 1:
@@ -307,6 +321,10 @@ def _sigmoid_topk_route(
     # Gather weights from original sigmoid scores (not bias-corrected)
     topk_weights = router_probs.gather(1, topk_indices)
 
+    # Accumulate counts for aux-free bias update
+    if afb_bias is not None:
+        _accumulate_afb_counts(moe_block, topk_indices)
+
     # Optional renormalization + scaling
     if getattr(moe_block, "norm_topk_prob", True):
         topk_weights = topk_weights / (topk_weights.sum(dim=-1, keepdim=True) + 1e-20)
@@ -335,6 +353,16 @@ def _route(moe_block, base_gate, hidden_states, gate_weight, gate_lora_delta):
     )
 
 
+def _accumulate_afb_counts(moe_block, topk_indices: torch.Tensor) -> None:
+    """Accumulate per-expert token counts for aux-free bias updates."""
+    afb_counts = getattr(moe_block, "_afb_counts", None)
+    if afb_counts is None:
+        return
+    flat_idx = topk_indices.reshape(-1)
+    counts = torch.bincount(flat_idx, minlength=afb_counts.numel())
+    afb_counts.add_(counts.to(afb_counts.dtype))
+
+
 # =============================================================================
 # Shared expert helpers
 # =============================================================================

src/axolotl/integrations/kernels/sonicmoe/routing.py

@@ -9,6 +9,12 @@ Different MoE architectures use different routing strategies:
 
 Each model type maps to a (routing_fn, activation_type, router_attr) triple.
 When routing_fn is None, the fused moe_TC_softmax_topk_layer path is used.
+
+Aux-loss-free (AFB) bias integration: when the aux_free_router plugin is
+active, ``moe_block._afb_bias`` and ``moe_block._afb_counts`` are registered
+as buffers.  The routing functions transparently inject the bias into expert
+*selection* (biased topk) while keeping mixture *weights* from unbiased
+scores, then accumulate per-expert token counts for the post-step bias update.
 """
 
 import torch
@@ -101,17 +107,25 @@ def softmax_topk_routing(
     router_logits = F.linear(hidden_states, gate.weight)  # [T, E]
     router_probs = F.softmax(router_logits, dim=-1, dtype=torch.float32)  # [T, E]
 
+    # Aux-free bias: biased selection, unbiased weights
+    afb_bias = getattr(moe_block, "_afb_bias", None)
+    scores_for_choice = router_probs
+    if afb_bias is not None:
+        scores_for_choice = router_probs + afb_bias
+
     # Select top-k experts per token
-    top_values, top_indices = torch.topk(router_probs, K, dim=-1)  # [T, K] each
+    top_values, top_indices = torch.topk(scores_for_choice, K, dim=-1)  # [T, K] each
+
+    # When aux-free bias is active, gather unbiased weights and accumulate counts
+    if afb_bias is not None:
+        top_values = router_probs.gather(1, top_indices)
+        _accumulate_afb_counts(moe_block, top_indices)
 
     # Renormalize if configured (default True for models without the attribute,
     # e.g. Mixtral/MiniMax which always normalize)
     if getattr(gate, "norm_topk_prob", True):
         top_values = top_values / top_values.sum(dim=-1, keepdim=True)
 
-    # no-op: matches transformers which casts to softmax output dtype (float32).
-    # top_values = top_values.to(router_probs.dtype)
-
     # Flatten for moe_general_routing_inputs.
     # Token indices are naturally sorted ascending from the [T, K] layout:
     # [0, 0, ..., 1, 1, ..., T-1, T-1, ...] — this is required by SonicMoE.
@@ -142,7 +156,11 @@ def softmax_group_topk_routing(
     router_logits = F.linear(hidden_states, gate.weight)  # [T, E]
     router_probs = F.softmax(router_logits, dim=-1, dtype=torch.float32)  # [T, E]
 
+    # Aux-free bias: inject before group selection / topk
+    afb_bias = getattr(moe_block, "_afb_bias", None)
     scores_for_choice = router_probs
+    if afb_bias is not None:
+        scores_for_choice = router_probs + afb_bias
 
     # Group selection: pick top groups, mask the rest
     if n_group > 1:
@@ -159,11 +177,17 @@ def softmax_group_topk_routing(
         score_mask = (
             group_mask.unsqueeze(-1).expand(-1, n_group, E // n_group).reshape(-1, E)
         )
-        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
+        scores_for_choice = scores_for_choice.masked_fill(
+            ~score_mask.bool(), -float("inf")
+        )
 
     topk_indices = torch.topk(scores_for_choice, k=K, dim=-1, sorted=False)[1]
     topk_weights = router_probs.gather(1, topk_indices)
 
+    # Accumulate counts for aux-free bias update
+    if afb_bias is not None:
+        _accumulate_afb_counts(moe_block, topk_indices)
+
     # Renormalization + scaling
     norm_topk_prob = getattr(moe_block, "norm_topk_prob", True)
     if norm_topk_prob:
@@ -233,6 +257,11 @@ def sigmoid_topk_routing(
         )
     scores_for_choice = router_probs + e_score_correction_bias
 
+    # Aux-free bias: stacks on top of e_score_correction_bias for selection
+    afb_bias = getattr(moe_block, "_afb_bias", None)
+    if afb_bias is not None:
+        scores_for_choice = scores_for_choice + afb_bias
+
     # Group-based selection: pick top groups, mask the rest (skip when n_group == 1)
     if n_group > 1:
         group_scores = (
@@ -248,7 +277,9 @@ def sigmoid_topk_routing(
         score_mask = (
             group_mask.unsqueeze(-1).expand(-1, n_group, E // n_group).reshape(-1, E)
         )
-        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
+        scores_for_choice = scores_for_choice.masked_fill(
+            ~score_mask.bool(), -float("inf")
+        )
 
     # Final topk from (possibly masked) scores
     topk_indices = torch.topk(scores_for_choice, k=K, dim=-1, sorted=False)[1]
@@ -256,6 +287,10 @@ def sigmoid_topk_routing(
     # Gather weights from original sigmoid scores (not bias-corrected)
     topk_weights = router_probs.gather(1, topk_indices)
 
+    # Accumulate counts for aux-free bias update
+    if afb_bias is not None:
+        _accumulate_afb_counts(moe_block, topk_indices)
+
     # Optional renormalization + scaling
     norm_topk_prob = getattr(moe_block, "norm_topk_prob", True)
     if norm_topk_prob:
@@ -276,3 +311,21 @@ def sigmoid_topk_routing(
     flat_expert_idx = topk_indices.to(torch.int32).reshape(-1)  # [T*K]
 
     return flat_scores, flat_token_idx, flat_expert_idx, router_logits
+
+
+def _accumulate_afb_counts(moe_block, topk_indices: torch.Tensor) -> None:
+    """Accumulate per-expert token counts for the aux-free bias update.
+
+    Called when ``moe_block._afb_bias`` is present (registered by the
+    ``aux_free_router`` plugin).  The counts are later consumed by the
+    ``MoeAuxFreeBiasUpdateCallback`` at each training step.
+    """
+    if hasattr(moe_block, "training") and not moe_block.training:
+        return
+    afb_counts = getattr(moe_block, "_afb_counts", None)
+    if afb_counts is None:
+        return
+    num_experts = afb_counts.numel()
+    flat_idx = topk_indices.reshape(-1)
+    counts = torch.bincount(flat_idx, minlength=num_experts)
+    afb_counts.add_(counts.to(afb_counts.dtype))

src/axolotl/utils/config/__init__.py

@@ -299,6 +299,7 @@ def validate_config(
     AxolotlInputConfig = AxolotlInputConfigBase
 
     if cfg.plugins:
+        prepare_plugins(cfg)
         (
             AxolotlConfigWCapabilities,
             AxolotlInputConfig,

src/axolotl/utils/schemas/config.py

@@ -758,44 +758,6 @@ class AxolotlInputConfig(
 
     llama4_linearized_experts: bool | None = None
 
-    # MoE aux-loss-free (AFB) toggles
-    moe_balance_type: Literal["gshard", "noaux_tc"] | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "MoE load balancing strategy: 'gshard' for auxiliary loss, 'noaux_tc' for aux-loss-free bias updates affecting top-k selection only. Defaults to model's native behavior when unset.",
-        },
-    )
-    moe_update_rate: float | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "Per-step bias update rate (gamma). Recommended: 0.005–0.05. If unset, plugin default is 0.01.",
-        },
-    )
-    moe_update_momentum: float | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "EMA momentum for expert load smoothing (0–1). If unset, plugin default is 0.9.",
-        },
-    )
-    moe_bias_cap: float | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "Absolute clamp for expert bias magnitude. If unset, plugin default is 2.0.",
-        },
-    )
-    moe_afb_warmup_steps: int | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "Number of initial steps to delay aux-free bias updates, allowing routing to stabilize. If unset, plugin default is 0.",
-        },
-    )
-    moe_bias_sync_group: Literal["world", "ep"] | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "Reduction group for expert load counts: 'world' (DP) or 'ep' (expert-parallel group if available). Defaults to 'world' when unset.",
-        },
-    )
-
     deepspeed: str | dict[str, Any] | None = Field(
         default=None,
         json_schema_extra={

tests/e2e/test_llama4_moe_aux_free.py

@@ -0,0 +1,75 @@
+"""
+E2E smoke test for Llama 4 aux-loss-free routing via plugin
+"""
+
+import unittest
+
+from axolotl.common.datasets import load_datasets
+from axolotl.train import train
+from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import check_model_output_exists, with_temp_dir
+
+
+class TestLlama4MoeAuxFree(unittest.TestCase):
+    """Smoke test to ensure aux-free plugin patches Llama 4 MoE correctly."""
+
+    @with_temp_dir
+    def test_llama4_aux_free_smoke(self, temp_dir):
+        cfg = DictDefault(
+            {
+                "base_model": "yujiepan/llama-4-tiny-random",
+                "tokenizer_config": "yujiepan/llama-4-tiny-random",
+                "trust_remote_code": False,
+                "flash_attention": False,
+                "sequence_len": 512,
+                "bf16": False,
+                "fp16": False,
+                "val_set_size": 0.02,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 1e-5,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "save_steps": 0,
+                "eval_steps": 0,
+                "save_first_step": False,
+                "plugins": [
+                    "axolotl.integrations.aux_free_router.plugin.AuxFreeMoEPlugin",
+                ],
+                "moe_balance_type": "noaux_tc",
+                "moe_update_rate": 0.01,
+                "moe_update_momentum": 0.9,
+                "moe_bias_cap": 2.0,
+            }
+        )
+
+        prepare_plugins(cfg)
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        dataset_meta = load_datasets(cfg=cfg)
+
+        model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
+
+        patched = next((m for m in model.modules() if hasattr(m, "_afb_bias")), None)
+        assert patched is not None, (
+            "Llama 4 MoE layer was not patched by aux-free plugin"
+        )
+        assert patched._afb_bias.ndim == 1
+        assert patched._afb_counts.ndim == 1
+        check_model_output_exists(temp_dir, cfg)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/e2e/test_moe_aux_free.py

@@ -8,7 +8,7 @@ import torch
 
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config, validate_config, prepare_plugins
+from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
 from axolotl.utils.dict import DictDefault
 
 from .utils import check_model_output_exists, with_temp_dir
@@ -57,9 +57,9 @@ class TestMoeAuxFree(unittest.TestCase):
             }
         )
 
+        prepare_plugins(cfg)
         cfg = validate_config(cfg)
         normalize_config(cfg)
-        prepare_plugins(cfg)
         dataset_meta = load_datasets(cfg=cfg)
 
         model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
@@ -67,7 +67,7 @@ class TestMoeAuxFree(unittest.TestCase):
         # Inspect model modules for a patched MoE layer
         patched = None
         for m in model.modules():
-            if hasattr(m, "_afb_patched") and getattr(m, "_afb_patched") is True:
+            if hasattr(m, "_afb_patched") and m._afb_patched is True:
                 patched = m
                 break
         assert patched is not None, "No MoE layer patched by aux-free plugin"

tests/e2e/test_moe_aux_parity.py

@@ -3,11 +3,14 @@ Parity test comparing aux-loss (gshard) vs aux-loss-free (noaux_tc) on Mixtral-t
 Checks that aux-free training loss does not degrade beyond a small tolerance.
 """
 
+import gc
 import unittest
 
+import torch
+
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config, validate_config, prepare_plugins
+from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
 from axolotl.utils.dict import DictDefault
 
 from .utils import with_temp_dir
@@ -61,6 +64,11 @@ class TestMoeAuxParity(unittest.TestCase):
         loss0 = _last_logged_loss(trainer0)
         assert loss0 is not None
 
+        # Release baseline resources before starting aux-free run
+        del model0, trainer0, dataset_meta0
+        gc.collect()
+        torch.cuda.empty_cache()
+
         # Aux-free: plugin + noaux_tc
         cfg1 = DictDefault(dict(base_cfg))
         cfg1.output_dir = f"{temp_dir}/auxfree"
@@ -71,9 +79,9 @@ class TestMoeAuxParity(unittest.TestCase):
         cfg1.moe_update_rate = 0.01
         cfg1.moe_update_momentum = 0.9
         cfg1.moe_bias_cap = 2.0
+        prepare_plugins(cfg1)
         cfg1 = validate_config(cfg1)
         normalize_config(cfg1)
-        prepare_plugins(cfg1)
         dataset_meta1 = load_datasets(cfg=cfg1)
         model1, _, trainer1 = train(cfg=cfg1, dataset_meta=dataset_meta1)
         loss1 = _last_logged_loss(trainer1)

tests/e2e/test_qwen3_moe_aux_free.py

@@ -4,11 +4,9 @@ E2E smoke test for Aux-Loss-Free MoE routing on Qwen3-MoE tiny
 
 import unittest
 
-import torch
-
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config, validate_config, prepare_plugins
+from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
 from axolotl.utils.dict import DictDefault
 
 from .utils import check_model_output_exists, with_temp_dir
@@ -55,9 +53,9 @@ class TestQwen3MoeAuxFree(unittest.TestCase):
             }
         )
 
+        prepare_plugins(cfg)
         cfg = validate_config(cfg)
         normalize_config(cfg)
-        prepare_plugins(cfg)
         dataset_meta = load_datasets(cfg=cfg)
 
         model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
@@ -65,7 +63,9 @@ class TestQwen3MoeAuxFree(unittest.TestCase):
         # check that at least one sparse MoE block has been patched
         found = False
         for m in model.modules():
-            if m.__class__.__name__.endswith("SparseMoeBlock") and hasattr(m, "_afb_patched"):
+            if m.__class__.__name__.endswith("SparseMoeBlock") and hasattr(
+                m, "_afb_patched"
+            ):
                 assert m._afb_patched is True
                 assert hasattr(m, "_afb_bias") and m._afb_bias.ndim == 1
                 assert hasattr(m, "_afb_counts") and m._afb_counts.ndim == 1

tests/e2e/test_ring_moe_aux_free.py

@@ -0,0 +1,74 @@
+"""
+E2E smoke test for Ring 2.0 aux-loss-free routing via plugin
+"""
+
+import unittest
+
+from axolotl.common.datasets import load_datasets
+from axolotl.train import train
+from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import check_model_output_exists, with_temp_dir
+
+
+class TestRingMoeAuxFree(unittest.TestCase):
+    """Smoke test to ensure aux-free plugin patches Ring Mini 2.0 correctly."""
+
+    @with_temp_dir
+    def test_ring_aux_free_smoke(self, temp_dir):
+        cfg = DictDefault(
+            {
+                "base_model": "yujiepan/ring-tiny-random",
+                "tokenizer_config": "yujiepan/ring-tiny-random",
+                "trust_remote_code": True,
+                "flash_attention": False,
+                "sequence_len": 512,
+                "bf16": False,
+                "fp16": False,
+                "val_set_size": 0.02,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 1e-5,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "save_steps": 0,
+                "eval_steps": 0,
+                "save_first_step": False,
+                # Aux-free plugin config
+                "plugins": [
+                    "axolotl.integrations.aux_free_router.plugin.AuxFreeMoEPlugin",
+                ],
+                "moe_balance_type": "noaux_tc",
+                "moe_update_rate": 0.01,
+                "moe_update_momentum": 0.9,
+                "moe_bias_cap": 2.0,
+            }
+        )
+
+        prepare_plugins(cfg)
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        dataset_meta = load_datasets(cfg=cfg)
+
+        model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
+
+        patched = next((m for m in model.modules() if hasattr(m, "_afb_bias")), None)
+        assert patched is not None, "Ring MoE layer was not patched by aux-free plugin"
+        assert patched._afb_bias.ndim == 1
+        assert patched._afb_counts.ndim == 1
+        check_model_output_exists(temp_dir, cfg)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/e2e/utils.py

@@ -12,6 +12,7 @@ from pathlib import Path
 
 import torch
 from packaging import version
+
 try:
     from tbparse import SummaryReader
 except ImportError:  # pragma: no cover - optional dependency
@@ -189,7 +190,9 @@ def check_tensorboard(
     helper function to parse and check tensorboard logs
     """
     if SummaryReader is None:
-        raise unittest.SkipTest("tbparse is not installed; skipping tensorboard assertions")
+        raise unittest.SkipTest(
+            "tbparse is not installed; skipping tensorboard assertions"
+        )
     tb_log_path = most_recent_subdir(temp_run_dir)
     event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
     reader = SummaryReader(event_file)

tests/unit/test_aux_free_adapters.py

@@ -2,14 +2,13 @@ import os
 import sys
 import tempfile
 import unittest
+from importlib import util as importlib_util
+from pathlib import Path
 from types import SimpleNamespace
 
 import torch
 import torch.distributed as dist
 import torch.nn as nn
-from importlib import util as importlib_util
-from pathlib import Path
-
 from huggingface_hub import snapshot_download
 
 from axolotl.integrations.aux_free_router.plugin import AuxFreeMoEPlugin
@@ -96,16 +95,21 @@ def _build_bailing_model():
 
 
 def _build_llama4_model():
-    from transformers import Llama4TextConfig
     from transformers.models.llama4.modeling_llama4 import Llama4TextMoe
 
-    config = Llama4TextConfig(
+    # Build config without __post_init__ validation (works around a
+    # huggingface_hub strict-dataclass type mismatch for layer_types).
+    config = object.__new__(__import__("transformers").Llama4TextConfig)
+    config.__dict__.update(
         hidden_size=16,
         intermediate_size=32,
         num_local_experts=4,
         num_attention_heads=2,
         num_key_value_heads=2,
         num_experts_per_tok=2,
+        num_hidden_layers=2,
+        hidden_act="silu",
+        layer_types=None,
     )
     layer = Llama4TextMoe(config)
 
@@ -148,12 +152,68 @@ def _build_mixtral_model():
     return DummyModel(layer), layer
 
 
-def _run_callback(plugin, cfg):
+def _build_qwen35_moe_model():
-    callbacks = plugin.add_callbacks_post_trainer(cfg, trainer=SimpleNamespace())
+    from transformers.models.qwen3_5_moe.configuration_qwen3_5_moe import (
+        Qwen3_5MoeTextConfig,
+    )
+    from transformers.models.qwen3_5_moe.modeling_qwen3_5_moe import (
+        Qwen3_5MoeSparseMoeBlock,
+    )
+
+    config = Qwen3_5MoeTextConfig(
+        hidden_size=16,
+        moe_intermediate_size=32,
+        shared_expert_intermediate_size=32,
+        num_experts=4,
+        num_experts_per_tok=2,
+        num_attention_heads=2,
+        num_key_value_heads=2,
+        num_hidden_layers=2,
+    )
+    layer = Qwen3_5MoeSparseMoeBlock(config)
+
+    class DummyModel(nn.Module):
+        def __init__(self, moe_layer):
+            super().__init__()
+            self.moe = moe_layer
+            self.config = SimpleNamespace(model_type="qwen3_5_moe")
+
+        def forward(self, hidden_states):
+            return self.moe(hidden_states)
+
+    return DummyModel(layer), layer
+
+
+def _run_callback(plugin, cfg, *, args=None, state=None, control=None):
+    if args is None:
+        args = SimpleNamespace(logging_steps=1)
+    if state is None:
+        state = SimpleNamespace(global_step=1, log_history=[])
+    if control is None:
+        control = SimpleNamespace(
+            should_log=False,
+            should_evaluate=False,
+            should_save=False,
+            should_training_stop=False,
+        )
+
+    class DummyTrainer:
+        def __init__(self, state_obj, control_obj):
+            self.state = state_obj
+            self.control = control_obj
+
+        def log(self, logs):
+            output = dict(logs)
+            output["step"] = self.state.global_step
+            self.state.log_history.append(output)
+            self.control.should_log = True
+
+    dummy_trainer = DummyTrainer(state, control)
+    callbacks = plugin.add_callbacks_post_trainer(cfg, trainer=dummy_trainer)
     assert callbacks, "expected aux-free callback to be registered"
     callback = callbacks[0]
-    dummy = SimpleNamespace()
+    callback.on_step_end(args=args, state=state, control=control)
-    callback.on_step_end(args=dummy, state=dummy, control=dummy)
+    return state, control
 
 
 class TestAuxFreeAdapters(unittest.TestCase):
@@ -170,7 +230,9 @@ class TestAuxFreeAdapters(unittest.TestCase):
 
         _run_callback(plugin, cfg)
         self.assertEqual(torch.count_nonzero(block._afb_counts), 0)
-        self.assertFalse(torch.allclose(block._afb_ema, torch.zeros_like(block._afb_ema)))
+        self.assertFalse(
+            torch.allclose(block._afb_ema, torch.zeros_like(block._afb_ema))
+        )
 
     def test_llama4_adapter_biases_router_selection(self):
         model, layer = _build_llama4_model()
@@ -185,7 +247,9 @@ class TestAuxFreeAdapters(unittest.TestCase):
 
         _run_callback(plugin, cfg)
         self.assertEqual(torch.count_nonzero(layer._afb_counts), 0)
-        self.assertFalse(torch.allclose(layer._afb_ema, torch.zeros_like(layer._afb_ema)))
+        self.assertFalse(
+            torch.allclose(layer._afb_ema, torch.zeros_like(layer._afb_ema))
+        )
 
     def test_bias_warmup_respected(self):
         model, block = _build_bailing_model()
@@ -193,48 +257,142 @@ class TestAuxFreeAdapters(unittest.TestCase):
         plugin = AuxFreeMoEPlugin()
         plugin.post_model_build(cfg, model)
 
-        callbacks = plugin.add_callbacks_post_trainer(cfg, trainer=SimpleNamespace())
-        self.assertTrue(callbacks)
-        callback = callbacks[0]
-        dummy = SimpleNamespace()
-
         def _step():
             hidden = torch.randn(2, 3, block.config.hidden_size)
             block(hidden)
-            callback.on_step_end(args=dummy, state=dummy, control=dummy)
+            _run_callback(plugin, cfg)
 
         # Warmup steps should leave bias untouched.
         _step()
-        self.assertTrue(torch.allclose(block._afb_bias, torch.zeros_like(block._afb_bias)))
+        self.assertTrue(
+            torch.allclose(block._afb_bias, torch.zeros_like(block._afb_bias))
+        )
 
         _step()
-        self.assertTrue(torch.allclose(block._afb_bias, torch.zeros_like(block._afb_bias)))
+        self.assertTrue(
+            torch.allclose(block._afb_bias, torch.zeros_like(block._afb_bias))
+        )
 
         # Third step exceeds warmup -> bias should update.
         _step()
         self.assertGreater(torch.count_nonzero(block._afb_bias), 0)
 
-    def test_mixtral_adapter_respects_native_forward(self):
+    def test_mixtral_adapter_patches_router_not_forward(self):
+        """Verify that aux-free patches the router (gate) only, and the
+        v5 block forward signature (single tensor return) is preserved."""
         model, layer = _build_mixtral_model()
-        layer.jitter_noise = 0.0  # avoid stochasticity for comparison
-
-        hidden_dim = layer.config.hidden_size
-        hidden = torch.randn(2, 3, hidden_dim)
-        baseline_out, baseline_logits = layer(hidden.clone())
-
         cfg = _cfg()
         plugin = AuxFreeMoEPlugin()
         plugin.post_model_build(cfg, model)
 
-        patched_out, patched_logits = layer(hidden.clone())
+        # Gate should be patched, not the block forward
-        self.assertTrue(torch.allclose(baseline_out, patched_out))
+        self.assertTrue(getattr(layer.gate, "_afb_patched", False))
-        self.assertTrue(torch.allclose(baseline_logits, patched_logits))
+        self.assertTrue(getattr(layer, "_afb_patched", False))
+
+        # v5 block forward returns a single tensor (not a tuple with logits)
+        hidden = torch.randn(2, 3, layer.config.hidden_size)
+        out = layer(hidden)
+        self.assertIsInstance(out, torch.Tensor)
+        self.assertEqual(out.shape, hidden.shape)
+
+        # Counts should have been accumulated
         self.assertGreater(torch.count_nonzero(layer._afb_counts), 0)
         _run_callback(plugin, cfg)
 
+    def test_mixtral_adapter_bias_affects_selection(self):
+        """When bias is large for one expert, it should be selected more often."""
+        model, layer = _build_mixtral_model()
+        cfg = _cfg()
+        plugin = AuxFreeMoEPlugin()
+        plugin.post_model_build(cfg, model)
+
+        # Set a large bias for expert 0 to force its selection
+        layer._afb_bias.zero_()
+        layer._afb_bias[0] = 10.0
+
+        hidden = torch.randn(2, 8, layer.config.hidden_size)
+        num_tokens = 2 * 8  # batch * seq
+        layer(hidden)
+
+        # With top_k=2, expert 0 should appear in every token's selection
+        # (once per token = num_tokens counts, not num_tokens * top_k)
+        counts = layer._afb_counts.clone()
+        self.assertEqual(
+            int(counts[0].item()),
+            num_tokens,
+            msg="Expert 0 should be selected for every token when heavily biased",
+        )
+
+    def test_qwen35_moe_adapter_patches_router_and_preserves_shared_expert(self):
+        """Verify Qwen 3.5 MoE: router is patched, shared expert is untouched,
+        output includes shared expert contribution."""
+        model, layer = _build_qwen35_moe_model()
+        cfg = _cfg()
+        plugin = AuxFreeMoEPlugin()
+        plugin.post_model_build(cfg, model)
+
+        # Gate should be patched
+        self.assertTrue(getattr(layer.gate, "_afb_patched", False))
+        self.assertTrue(getattr(layer, "_afb_patched", False))
+        # Shared expert should be unmodified
+        self.assertTrue(hasattr(layer, "shared_expert"))
+        self.assertTrue(hasattr(layer, "shared_expert_gate"))
+
+        # Forward should return a single tensor (shared + routed)
+        hidden_size = layer.gate.hidden_dim
+        hidden = torch.randn(2, 3, hidden_size)
+        out = layer(hidden)
+        self.assertIsInstance(out, torch.Tensor)
+        self.assertEqual(out.shape, hidden.shape)
+
+        # Counts should have been accumulated
+        self.assertGreater(torch.count_nonzero(layer._afb_counts), 0)
+
+    def test_qwen35_moe_adapter_bias_updates(self):
+        """Full cycle: forward → callback → verify bias update for Qwen 3.5 MoE."""
+        model, layer = _build_qwen35_moe_model()
+        cfg = _cfg()
+        plugin = AuxFreeMoEPlugin()
+        plugin.post_model_build(cfg, model)
+
+        hidden_size = layer.gate.hidden_dim
+        hidden = torch.randn(2, 4, hidden_size)
+        layer(hidden)
+
+        # Bias should start at zero
+        self.assertTrue(
+            torch.allclose(layer._afb_bias, torch.zeros_like(layer._afb_bias))
+        )
+
+        _run_callback(plugin, cfg)
+
+        # After callback: counts reset, EMA updated, bias updated
+        self.assertEqual(torch.count_nonzero(layer._afb_counts), 0)
+        self.assertFalse(
+            torch.allclose(layer._afb_ema, torch.zeros_like(layer._afb_ema))
+        )
+
+    def test_qwen35_moe_adapter_model_type_matching(self):
+        """Verify the adapter matches both qwen3_5_moe and qwen3_5_moe_text."""
+        from axolotl.integrations.aux_free_router.adapters import Qwen35MoeAdapter
+
+        adapter = Qwen35MoeAdapter()
+
+        model_moe = SimpleNamespace(config=SimpleNamespace(model_type="qwen3_5_moe"))
+        model_text = SimpleNamespace(
+            config=SimpleNamespace(model_type="qwen3_5_moe_text")
+        )
+        model_other = SimpleNamespace(config=SimpleNamespace(model_type="qwen3_moe"))
+
+        self.assertTrue(adapter.matches(model_moe))
+        self.assertTrue(adapter.matches(model_text))
+        self.assertFalse(adapter.matches(model_other))
+
     def test_ep_group_resolution_deferred_until_dist_ready(self):
         if dist.is_available() and dist.is_initialized():
-            dist.destroy_process_group()
+            self.skipTest(
+                "Cannot safely test deferred EP group resolution when a process group is already initialized"
+            )
 
         model, block = _build_bailing_model()
         cfg = _cfg(moe_bias_sync_group="ep", expert_parallel_size=1)
@@ -244,24 +402,265 @@ class TestAuxFreeAdapters(unittest.TestCase):
         self.assertIsNotNone(plugin._shim)
         self.assertIsNone(plugin._shim.ep_group)
 
-        callbacks = plugin.add_callbacks_post_trainer(cfg, trainer=SimpleNamespace())
-        self.assertTrue(callbacks)
-        callback = callbacks[0]
-        dummy = SimpleNamespace()
-
         tmp_init = tempfile.NamedTemporaryFile(delete=False)
         tmp_init.close()
         init_method = f"file://{tmp_init.name}"
-        dist.init_process_group(backend="gloo", init_method=init_method, world_size=1, rank=0)
+        dist.init_process_group(
+            backend="gloo", init_method=init_method, world_size=1, rank=0
+        )
         try:
             hidden = torch.randn(2, 3, block.config.hidden_size)
             block(hidden)
-            callback.on_step_end(args=dummy, state=dummy, control=dummy)
+            _run_callback(
+                plugin,
+                cfg,
+                args=SimpleNamespace(logging_steps=1),
+                state=SimpleNamespace(global_step=1, log_history=[]),
+                control=SimpleNamespace(
+                    should_log=False,
+                    should_evaluate=False,
+                    should_save=False,
+                    should_training_stop=False,
+                ),
+            )
             self.assertIs(plugin._shim.ep_group, dist.group.WORLD)
         finally:
             dist.destroy_process_group()
             os.unlink(tmp_init.name)
 
+    def test_telemetry_logging(self):
+        model, layer = _build_mixtral_model()
+        cfg = _cfg()
+        plugin = AuxFreeMoEPlugin()
+        plugin.post_model_build(cfg, model)
+
+        hidden_dim = layer.config.hidden_size
+        hidden = torch.randn(2, 3, hidden_dim)
+        layer(hidden)
+
+        args = SimpleNamespace(logging_steps=1)
+        state = SimpleNamespace(global_step=1, log_history=[])
+        control = SimpleNamespace(
+            should_log=False,
+            should_evaluate=False,
+            should_save=False,
+            should_training_stop=False,
+        )
+        _run_callback(plugin, cfg, args=args, state=state, control=control)
+
+        self.assertTrue(control.should_log)
+        self.assertTrue(state.log_history)
+        telemetry = state.log_history[-1]
+        self.assertEqual(telemetry["step"], state.global_step)
+        self.assertIn("moe_afb/l0_load_min", telemetry)
+        self.assertIn("moe_afb/l0_load_max", telemetry)
+        self.assertIn("moe_afb/l0_bias_abs_max", telemetry)
+
+    def test_get_num_experts_v5_attribute_paths(self):
+        """Verify get_num_experts works with v5 attribute layout where
+        num_experts is on gate/experts sub-modules, not the block."""
+        from axolotl.integrations.aux_free_router.adapters import MixtralAdapter
+
+        adapter = MixtralAdapter()
+
+        # Simulates v5 MixtralSparseMoeBlock (num_experts on gate, not block)
+        block = SimpleNamespace(
+            gate=SimpleNamespace(num_experts=8),
+            experts=SimpleNamespace(num_experts=8),
+        )
+        self.assertEqual(adapter.get_num_experts(block), 8)
+
+        # Also works when num_experts is directly on block
+        block2 = SimpleNamespace(num_experts=4)
+        self.assertEqual(adapter.get_num_experts(block2), 4)
+
+
+class TestAuxFreeKernelComposition(unittest.TestCase):
+    """Tests that aux-free bias composes correctly with kernel routing."""
+
+    def test_sonicmoe_softmax_routing_with_afb_bias(self):
+        """SonicMoE softmax routing should use biased selection / unbiased weights."""
+        from axolotl.integrations.kernels.sonicmoe.routing import softmax_topk_routing
+
+        num_experts = 4
+        top_k = 2
+        hidden_dim = 16
+        T = 6
+
+        # Build a mock MoE block with gate attributes
+        gate = nn.Linear(hidden_dim, num_experts, bias=False)
+        gate.top_k = top_k
+        gate.num_experts = num_experts
+        gate.norm_topk_prob = True
+
+        moe_block = SimpleNamespace(gate=gate)
+        hidden = torch.randn(T, hidden_dim)
+
+        # Baseline: no bias
+        scores_base, tok_base, exp_base, logits_base = softmax_topk_routing(
+            hidden, moe_block
+        )
+        self.assertEqual(scores_base.shape[0], T * top_k)
+
+        # Now register aux-free buffers and set heavy bias on expert 0
+        moe_block._afb_bias = torch.zeros(num_experts)
+        moe_block._afb_bias[0] = 100.0
+        moe_block._afb_counts = torch.zeros(num_experts)
+
+        scores_biased, tok_biased, exp_biased, logits_biased = softmax_topk_routing(
+            hidden, moe_block
+        )
+
+        # Expert 0 should be selected for every token
+        self.assertTrue(
+            (exp_biased == 0).any(),
+            "Expert 0 should appear in selections when heavily biased",
+        )
+        # Counts should have been accumulated
+        self.assertGreater(moe_block._afb_counts[0].item(), 0)
+        # Total counts should equal T * top_k
+        self.assertEqual(int(moe_block._afb_counts.sum().item()), T * top_k)
+
+    def test_sonicmoe_routing_without_bias_unchanged(self):
+        """Without _afb_bias, routing should produce identical results."""
+        from axolotl.integrations.kernels.sonicmoe.routing import softmax_topk_routing
+
+        num_experts = 4
+        top_k = 2
+        hidden_dim = 16
+
+        gate = nn.Linear(hidden_dim, num_experts, bias=False)
+        gate.top_k = top_k
+        gate.num_experts = num_experts
+        gate.norm_topk_prob = True
+
+        moe_block = SimpleNamespace(gate=gate)
+        hidden = torch.randn(4, hidden_dim)
+
+        # Without _afb_bias attribute
+        scores1, _, exp1, _ = softmax_topk_routing(hidden, moe_block)
+
+        # With _afb_bias = zeros (should be equivalent)
+        moe_block._afb_bias = torch.zeros(num_experts)
+        moe_block._afb_counts = torch.zeros(num_experts)
+        scores2, _, exp2, _ = softmax_topk_routing(hidden, moe_block)
+
+        torch.testing.assert_close(scores1, scores2)
+        torch.testing.assert_close(exp1, exp2)
+
+    @unittest.skipUnless(
+        importlib_util.find_spec("triton") is not None,
+        "triton not installed (required by scattermoe)",
+    )
+    def test_scattermoe_softmax_routing_with_afb_bias(self):
+        """ScatterMoE softmax routing should use biased selection / unbiased weights."""
+        from axolotl.integrations.kernels.libs.scattermoe_lora.layers import (
+            _softmax_topk_route,
+        )
+
+        num_experts = 4
+        top_k = 2
+        hidden_dim = 16
+        T = 6
+
+        gate_weight = torch.randn(num_experts, hidden_dim)
+        base_gate = SimpleNamespace(
+            top_k=top_k,
+            num_experts=num_experts,
+            norm_topk_prob=True,
+            weight=gate_weight,
+        )
+
+        moe_block = SimpleNamespace()
+        hidden = torch.randn(T, hidden_dim)
+
+        # Baseline without bias
+        w_base, e_base, _, _ = _softmax_topk_route(
+            moe_block, base_gate, hidden, gate_weight, None
+        )
+
+        # With heavy bias on expert 0
+        moe_block._afb_bias = torch.zeros(num_experts)
+        moe_block._afb_bias[0] = 100.0
+        moe_block._afb_counts = torch.zeros(num_experts)
+
+        w_biased, e_biased, _, _ = _softmax_topk_route(
+            moe_block, base_gate, hidden, gate_weight, None
+        )
+
+        # Expert 0 should appear in all selections
+        self.assertTrue((e_biased == 0).any())
+        # Counts accumulated
+        self.assertGreater(moe_block._afb_counts[0].item(), 0)
+        self.assertEqual(int(moe_block._afb_counts.sum().item()), T * top_k)
+
+    def test_kernel_routing_skips_router_patch(self):
+        """When a kernel backend has patched the block class, the adapter
+        should skip patching the router (buffers are still registered)."""
+        from axolotl.integrations.aux_free_router.adapters import MixtralAdapter
+
+        adapter = MixtralAdapter()
+
+        # Create a mock layer whose class has _original_forward (SonicMoE marker)
+        class PatchedBlock(nn.Module):
+            _original_forward = True  # SonicMoE marker
+
+            def __init__(self):
+                super().__init__()
+                self.gate = nn.Linear(16, 4, bias=False)
+                self.gate.top_k = 2
+                self.gate.num_experts = 4
+                self.gate.hidden_dim = 16
+                self.experts = nn.Linear(16, 16)  # placeholder
+
+        layer = PatchedBlock()
+        self.assertTrue(adapter.uses_kernel_routing(layer))
+
+        # Gate should NOT be patched (kernel handles routing)
+        self.assertFalse(getattr(layer.gate, "_afb_patched", False))
+
+    def test_adapter_buffers_registered_even_with_kernel(self):
+        """Even when kernel routing is active, aux-free buffers must be
+        registered on the MoE block so the kernel routing can find them."""
+        from axolotl.integrations.aux_free_router.adapters import (
+            LayerHandle,
+            MixtralAdapter,
+        )
+        from axolotl.integrations.aux_free_router.core import (
+            AuxFreeConfig,
+            AuxFreeShim,
+            AuxFreeState,
+        )
+
+        class PatchedBlock(nn.Module):
+            _original_forward = True
+
+            def __init__(self):
+                super().__init__()
+                self.gate = nn.Linear(16, 4, bias=False)
+                self.gate.top_k = 2
+                self.gate.num_experts = 4
+                self.gate.hidden_dim = 16
+                self.experts = nn.Linear(16, 16)
+
+        layer = PatchedBlock()
+        adapter = MixtralAdapter()
+        cfg = AuxFreeConfig()
+        state = AuxFreeState(
+            num_layers=1, num_experts=4, device=torch.device("cpu"), cfg=cfg
+        )
+        shim = AuxFreeShim(state=state)
+        handle = LayerHandle(layer=layer, layer_idx=0, num_experts=4, top_k=2)
+
+        adapter.prepare(layer, handle, shim)
+
+        # Buffers should be registered for kernel routing to use
+        self.assertTrue(hasattr(layer, "_afb_bias"))
+        self.assertTrue(hasattr(layer, "_afb_counts"))
+        self.assertTrue(hasattr(layer, "_afb_ema"))
+        # But gate should NOT be patched
+        self.assertFalse(getattr(layer.gate, "_afb_patched", False))
+
 
 if __name__ == "__main__":
     unittest.main()