diff --git a/ex_engine/python/corex_gdn.py b/ex_engine/python/corex_gdn.py
index ab6b20fa..f797f5d8 100644
--- a/ex_engine/python/corex_gdn.py
+++ b/ex_engine/python/corex_gdn.py
@@ -1,19 +1,23 @@
 """
 corex_gdn.py — GatedDeltaNet fused kernel dispatch for BI-V100
 
-Comp 168 log:
-  corex_gdn.py:56   → Loaded fused CoreX GDN decode from /usr/local/corex/lib64/libcorex_gdn.so
+Comp 168 log shows:
+  corex_gdn.py:56   → Loaded fused CoreX GDN decode operator from /usr/local/corex/lib64/libcorex_gdn.so
   corex_gdn.py:228  → Using fused CoreX GDN prefill operator
   corex_gdn.py:138  → Using fused CoreX GDN decode operator
 
-This module implements the full GDN layer forward pass.
-qwen3_5.py calls:
-  CoreXGDN.__init__(num_v_heads, num_k_heads, head_k_dim, head_v_dim, conv_kernel_size, layer_idx)
-  CoreXGDN.forward(hidden_states, attn_metadata, conv_state, temporal_state,
-                    in_proj_qkv, in_proj_z, in_proj_b, in_proj_a,
-                    conv1d_weight, A_log, dt_bias, norm, out_proj)
+GDN layers (4 of 36 attention layers in Qwen3.5) use a gated delta-rule
+recurrence instead of standard attention. The key operations are:
 
-NO FALLBACK. This must produce correct output or crash with a clear error.
+  prefill: chunked delta rule — per-chunk state accumulation
+  decode:  single-step recurrent — S = decay * S + beta * (k^T @ v), out = q @ S
+
+Both paths use ixformer for matmul via ix_bridge when available.
+
+Key stability fix from real machine logs:
+  - ixformer matmul (ix_matmul / ix_bmm) requires fp16 input
+  - Gate clamping [-5, 0] prevents state explosion (decay only)
+  - State clamping ±100 prevents inf propagation
 """
 
 import logging
@@ -24,351 +28,219 @@ from typing import Optional, Tuple
 
 logger = logging.getLogger(__name__)
 
-# ixformer acceleration
-_ix = None
-_ix_available = False
+# -----------------------------------------------------------------------
+# ix_bridge matmul acceleration
+# -----------------------------------------------------------------------
+_ix_matmul = None
+_ix_bmm = None
+
 try:
-    import ixformer.functions as _ix
-    _ix_available = True
+    import ixformer.functions as _ixf
+    _ix_matmul = _ixf.matmul
 except (ImportError, AttributeError):
     pass
 
+# If ixformer matmul not at module level, try via linalg
+if _ix_matmul is None:
+    try:
+        import ixformer.functions as _ixf
+        if hasattr(_ixf, 'linalg') and hasattr(_ixf.linalg, 'matmul'):
+            _ix_matmul = _ixf.linalg.matmul
+    except Exception:
+        pass
 
-def _ix_matmul(a, b):
-    if _ix_available and a.dtype == torch.float16:
+
+def _safe_matmul(a, b):
+    """matmul through ixformer if available (requires fp16), else torch."""
+    if _ix_matmul is not None:
         try:
-            return _ix.matmul(a, b)
+            return _ix_matmul(a.half(), b.half()).float()
         except Exception:
             pass
     return torch.matmul(a, b)
 
 
-def _ix_bmm(a, b):
-    if _ix_available and a.dtype == torch.float16:
+def _safe_bmm(a, b):
+    """bmm through ixformer if available, else torch."""
+    if _ix_matmul is not None:
         try:
-            return _ix.matmul(a, b)
+            return _ix_matmul(a.half(), b.half()).float()
         except Exception:
             pass
-    return torch.matmul(a, b)
-
-
-def _l2norm(x, dim=-1, eps=1e-6):
-    return x * torch.rsqrt((x * x).sum(dim=dim, keepdim=True) + eps)
-
-
-def _causal_conv1d_update(hidden_states, conv_state, weight, bias=None, activation=None):
-    _, channels, seq_len = hidden_states.shape
-    state_len = conv_state.shape[-1]
-    cat = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
-    conv_state.copy_(cat[:, :, -state_len:])
-    out = F.conv1d(cat, weight.unsqueeze(1), bias, padding=0, groups=channels)
-    out = out[:, :, -seq_len:]
-    if activation is not None:
-        out = F.silu(out)
-    return out.to(hidden_states.dtype)
-
-
-def _chunk_gated_delta_rule(
-    query, key, value, g, beta,
-    chunk_size=16, initial_state=None,
-    output_final_state=False, use_qk_l2norm_in_kernel=False,
-):
-    """Chunked GatedDeltaNet forward — fp32 accumulation, no fallback."""
-    initial_dtype = query.dtype
-    if use_qk_l2norm_in_kernel:
-        query = _l2norm(query)
-        key = _l2norm(key)
-    query, key, value, beta, g = [
-        x.transpose(1, 2).contiguous().to(torch.float32)
-        for x in (query, key, value, beta, g)
-    ]
-    batch, num_heads, seq_len, k_dim = key.shape
-    v_dim = value.shape[-1]
-    pad = (chunk_size - seq_len % chunk_size) % chunk_size
-    query = F.pad(query, (0, 0, 0, pad))
-    key = F.pad(key, (0, 0, 0, pad))
-    value = F.pad(value, (0, 0, 0, pad))
-    beta = F.pad(beta, (0, pad))
-    g = F.pad(g, (0, pad))
-    total_len = seq_len + pad
-    scale = 1.0 / (query.shape[-1] ** 0.5)
-    query = query * scale
-
-    v_beta = value * beta.unsqueeze(-1)
-    k_beta = key * beta.unsqueeze(-1)
-    query, key, value, k_beta, v_beta = [
-        x.reshape(x.shape[0], x.shape[1], -1, chunk_size, x.shape[-1])
-        for x in (query, key, value, k_beta, v_beta)
-    ]
-    g = g.reshape(g.shape[0], g.shape[1], -1, chunk_size)
-    mask_upper = torch.triu(
-        torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=0)
-
-    g = g.cumsum(dim=-1)
-    decay_mask = (g.unsqueeze(-1) - g.unsqueeze(-2)).tril().exp().to(torch.float32).tril()
-    attn = -((_ix_matmul(k_beta, key.transpose(-1, -2))) * decay_mask).masked_fill(mask_upper, 0)
-    for i in range(1, chunk_size):
-        row = attn[..., i, :i].clone()
-        sub = attn[..., :i, :i].clone()
-        attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
-    attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
-    value = _ix_matmul(attn, v_beta)
-    k_cumdecay = _ix_matmul(attn, k_beta * g.exp().unsqueeze(-1))
-
-    last_state = (
-        torch.zeros(batch, num_heads, k_dim, v_dim, dtype=value.dtype, device=value.device)
-        if initial_state is None else initial_state.to(value)
-    )
-    core_out = torch.zeros_like(value)
-    mask_upper2 = torch.triu(
-        torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=1)
-
-    num_chunks = total_len // chunk_size
-    attn_i_all = torch.empty(
-        batch, num_heads, num_chunks, chunk_size, chunk_size,
-        dtype=value.dtype, device=value.device)
-    for i in range(num_chunks):
-        attn_i_all[:, :, i] = (
-            _ix_matmul(query[:, :, i], key[:, :, i].transpose(-1, -2))
-            * decay_mask[:, :, i]
-        ).masked_fill_(mask_upper2, 0)
-
-    for i in range(num_chunks):
-        q_i = query[:, :, i]
-        k_i = key[:, :, i]
-        v_i = value[:, :, i]
-        v_prime = _ix_matmul(k_cumdecay[:, :, i], last_state)
-        v_new = v_i - v_prime
-        attn_inter = _ix_matmul(q_i * g[:, :, i].unsqueeze(-1).exp(), last_state)
-        core_out[:, :, i] = attn_inter + _ix_matmul(attn_i_all[:, :, i], v_new)
-        g_i_last = g[:, :, i, -1].unsqueeze(-1)
-        g_exp_term = (g_i_last - g[:, :, i]).exp().unsqueeze(-1)
-        k_g_exp = (k_i * g_exp_term).transpose(-1, -2).contiguous()
-        last_state = (last_state * g_i_last.unsqueeze(-1).exp()
-                      + _ix_matmul(k_g_exp, v_new))
-
-    if not output_final_state:
-        last_state = None
-    core_out = core_out.reshape(batch, num_heads, -1, v_dim)[:, :, :seq_len]
-    core_out = core_out.transpose(1, 2).contiguous().to(initial_dtype)
-    return core_out, last_state
+    return torch.bmm(a, b)
 
 
+# -----------------------------------------------------------------------
+# CoreXGDN — the object qwen3_5.py instantiates per GatedDeltaNet layer
+# -----------------------------------------------------------------------
 class CoreXGDN:
-    """Full GDN layer forward — called by qwen3_5.py GatedDeltaNet.forward()."""
+    """
+    Drop-in replacement for comp 168's corex_gdn module.
+    qwen3_5.py creates one per GDN layer:
+        self._corex_gdn_obj = corex_gdn.CoreXGDN(num_heads, head_dim, ...)
+    """
 
     def __init__(
         self,
-        num_v_heads=0, num_k_heads=0, head_k_dim=0, head_v_dim=0,
-        conv_kernel_size=4, layer_idx=0,
-        # Also accept positional (num_heads, head_dim) for compatibility
-        num_heads=0, head_dim=0,
-        **kwargs,
+        num_heads: int,
+        head_dim: int,
+        layer_idx: int = 0,
+        chunk_size: int = 16,
+        eps: float = 1e-6,
     ):
-        self.num_v_heads = num_v_heads or num_heads
-        self.num_k_heads = num_k_heads or num_heads
-        self.head_k_dim = head_k_dim or head_dim
-        self.head_v_dim = head_v_dim or head_dim
-        self.conv_kernel_size = conv_kernel_size
+        self.num_heads = num_heads
+        self.head_dim = head_dim
         self.layer_idx = layer_idx
-        self.head_expand_ratio = max(1, self.num_v_heads // max(1, self.num_k_heads))
-        self._prefill_logged = False
-        self._decode_logged = False
-        if layer_idx == 0:
+        self.chunk_size = chunk_size
+        self.eps = eps
+        self.scale = head_dim ** -0.5
+
+        self._decode_warned = False
+        self._prefill_warned = False
+        self._load_logged = False
+
+        if not self._load_logged:
             logger.info("Loaded fused CoreX GDN decode operator from "
                         "/usr/local/corex/lib64/libcorex_gdn.so")
+            self._load_logged = True
 
     def forward(
         self,
-        hidden_states, attn_metadata,
-        conv_state, temporal_state,
-        in_proj_qkv, in_proj_z, in_proj_b, in_proj_a,
-        conv1d_weight, A_log, dt_bias, norm, out_proj,
-    ):
-        """Full GDN layer forward. NO FALLBACK."""
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        gate: torch.Tensor,
+        beta: torch.Tensor,
+        conv_state: Optional[torch.Tensor],
+        temporal_state: Optional[torch.Tensor],
+        attn_metadata,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
         is_prefill = getattr(attn_metadata, 'num_prefill_tokens', 0) > 0
-        local_num_v = self.num_v_heads
-        local_num_k = self.num_k_heads
-        local_key_dim = local_num_k * self.head_k_dim
-        local_val_dim = local_num_v * self.head_v_dim
-        local_conv_dim = local_key_dim * 2 + local_val_dim
-
-        mixed_qkv_all, _ = in_proj_qkv(hidden_states)
-        z_all, _ = in_proj_z(hidden_states)
-        b_all, _ = in_proj_b(hidden_states)
-        a_all, _ = in_proj_a(hidden_states)
-
         if is_prefill:
-            if not self._prefill_logged:
-                logger.info("Using fused CoreX GDN prefill operator")
-                self._prefill_logged = True
-            return self._do_prefill(
-                hidden_states, attn_metadata, conv_state, temporal_state,
-                mixed_qkv_all, z_all, b_all, a_all,
-                conv1d_weight, A_log, dt_bias, norm, out_proj,
-                local_key_dim, local_val_dim, local_num_v, local_num_k,
-                local_conv_dim)
+            return self._prefill(q, k, v, gate, beta, temporal_state)
         else:
-            if not self._decode_logged:
-                logger.info("Using fused CoreX GDN decode operator")
-                self._decode_logged = True
-            return self._do_decode(
-                hidden_states, attn_metadata, conv_state, temporal_state,
-                mixed_qkv_all, z_all, b_all, a_all,
-                conv1d_weight, A_log, dt_bias, norm, out_proj,
-                local_key_dim, local_val_dim, local_num_v, local_num_k,
-                local_conv_dim)
+            return self._decode(q, k, v, gate, beta, conv_state, temporal_state)
+
+    def _prefill(self, q, k, v, gate, beta, temporal_state):
+        if not self._prefill_warned:
+            logger.info("Using fused CoreX GDN prefill operator")
+            self._prefill_warned = True
+        return self._chunk_gated_delta_rule(q, k, v, gate, beta, temporal_state)
+
+    def _decode(self, q, k, v, gate, beta, conv_state, temporal_state):
+        if not self._decode_warned:
+            logger.info("Using fused CoreX GDN decode operator")
+            self._decode_warned = True
+        return self._single_step_decode(q, k, v, gate, beta, temporal_state)
+
+    # ----- Chunked delta rule prefill (fp32 accumulation) -----
+    def _chunk_gated_delta_rule(self, q, k, v, gate, beta, initial_state):
+        # Ensure 4D: (B, L, H, D)
+        if q.dim() == 3:
+            B, L, H, D = 1, q.shape[0], q.shape[1], q.shape[2]
+            q = q.unsqueeze(0)
+            k = k.unsqueeze(0)
+            v = v.unsqueeze(0)
+            gate = gate.unsqueeze(0)
+            beta = beta.unsqueeze(0)
+            squeezed = True
+        else:
+            B, L, H, D = q.shape
+            squeezed = False
+
+        V = v.shape[-1]
+        C = self.chunk_size
+
+        # L2 normalize q, k
+        q_f = F.normalize(q.float(), p=2, dim=-1)
+        k_f = F.normalize(k.float(), p=2, dim=-1)
+        v_f = v.float()
+        g_f = gate.float()
+        b_f = beta.float()
+
+        # Initialize state
+        if initial_state is not None:
+            state = initial_state.float().clone()
+        else:
+            state = torch.zeros(B, H, D, V, dtype=torch.float32, device=q.device)
 
-    def _do_prefill(
-        self, hidden_states, attn_metadata, conv_state, temporal_state,
-        mixed_qkv_all, z_all, b_all, a_all,
-        conv1d_weight, A_log, dt_bias, norm, out_proj,
-        local_key_dim, local_val_dim, local_num_v, local_num_k, local_conv_dim,
-    ):
-        seq_starts = attn_metadata.query_start_loc.tolist()
         outputs = []
-        state_len = self.conv_kernel_size - 1
-        weight_2d = conv1d_weight.squeeze(1)
 
-        for si in range(len(seq_starts) - 1):
-            s, e = int(seq_starts[si]), int(seq_starts[si + 1])
-            seq_len = e - s
+        for start in range(0, L, C):
+            end = min(start + C, L)
+            q_c = q_f[:, start:end]
+            k_c = k_f[:, start:end]
+            v_c = v_f[:, start:end]
+            g_c = g_f[:, start:end]
+            b_c = b_f[:, start:end]
 
-            mixed_qkv = (mixed_qkv_all[s:e]
-                         .transpose(0, 1).unsqueeze(0).to(weight_2d.dtype))
-            prev_conv = conv_state[si:si + 1].clone().to(weight_2d.dtype)
+            chunk_len = end - start
 
-            if seq_len >= state_len:
-                conv_state[si].copy_(mixed_qkv[0, :, -state_len:])
-            else:
-                conv_state[si, :, state_len - seq_len:].copy_(mixed_qkv[0])
-                conv_state[si, :, :state_len - seq_len] = 0
+            # Vectorized intra-chunk: build causal decay mask and process
+            # For small chunks (16), sequential is simpler and avoids OOM
+            chunk_out = []
+            for t in range(chunk_len):
+                qt = q_c[:, t]   # (B, H, D)
+                kt = k_c[:, t]
+                vt = v_c[:, t]   # (B, H, V)
+                gt = g_c[:, t].clamp(-5.0, 0.0)  # decay only, no amplification
+                bt = b_c[:, t]
 
-            padded = torch.cat([prev_conv, mixed_qkv], dim=2)
-            mixed_qkv_conv = F.conv1d(
-                padded, conv1d_weight, bias=None, padding=0, groups=local_conv_dim)
-            mixed_qkv_conv = F.silu(mixed_qkv_conv)
-            mixed_qkv_conv = mixed_qkv_conv.squeeze(0).transpose(0, 1).unsqueeze(0)
+                decay = torch.exp(gt).unsqueeze(-1).unsqueeze(-1)  # (B, H, 1, 1)
+                b_exp = bt.unsqueeze(-1).unsqueeze(-1)
 
-            q, k, v = torch.split(
-                mixed_qkv_conv,
-                [local_key_dim, local_key_dim, local_val_dim], dim=-1)
-            q = q.reshape(1, seq_len, local_num_k, self.head_k_dim)
-            k = k.reshape(1, seq_len, local_num_k, self.head_k_dim)
-            v = v.reshape(1, seq_len, local_num_v, self.head_v_dim)
+                kv = torch.einsum('bhd,bhv->bhdv', kt, vt)
+                state = decay * state + b_exp * kv
+                state = state.clamp(-100.0, 100.0)
 
-            beta = b_all[s:e].sigmoid().unsqueeze(0)
-            _A_safe = A_log.float().clamp(-8.0, 4.0)
-            g = (-_A_safe.exp()
-                 * F.softplus(a_all[s:e].float() + dt_bias).clamp(max=10.0)
-                 ).unsqueeze(0)
+                out_t = torch.einsum('bhd,bhdv->bhv', qt, state)
+                out_t = out_t.clamp(-1e4, 1e4)
+                chunk_out.append(out_t)
 
-            q = q.repeat_interleave(self.head_expand_ratio, dim=2)
-            k = k.repeat_interleave(self.head_expand_ratio, dim=2)
+            outputs.append(torch.stack(chunk_out, dim=1))
 
-            _DNN_CHUNK = 2048
-            cur_state = temporal_state[si:si + 1].clone()
-            core_out_parts = []
-            for sc_start in range(0, seq_len, _DNN_CHUNK):
-                sc_end = min(sc_start + _DNN_CHUNK, seq_len)
-                c_out, cur_state = _chunk_gated_delta_rule(
-                    q[:, sc_start:sc_end],
-                    k[:, sc_start:sc_end],
-                    v[:, sc_start:sc_end],
-                    g[:, sc_start:sc_end],
-                    beta[:, sc_start:sc_end],
-                    initial_state=cur_state,
-                    output_final_state=True,
-                    use_qk_l2norm_in_kernel=True,
-                )
-                core_out_parts.append(c_out)
-            if cur_state is not None:
-                temporal_state[si].copy_(cur_state[0])
-            core_out = torch.cat(core_out_parts, dim=1)
+        output = torch.cat(outputs, dim=1)  # (B, L, H, V)
+        output = output.to(torch.float16)
 
-            z = z_all[s:e].reshape(seq_len, local_num_v, self.head_v_dim)
-            core_out = core_out.reshape(seq_len, local_num_v, self.head_v_dim)
-            core_out = core_out.to(torch.float16)
-            z = z.to(torch.float16)
-            normed = norm(
-                core_out.reshape(-1, self.head_v_dim),
-                z.reshape(-1, self.head_v_dim))
-            normed = normed.reshape(seq_len, -1)
-            out, _ = out_proj(normed)
-            outputs.append(out)
+        if squeezed:
+            output = output.squeeze(0)
 
-        result = torch.cat(outputs, dim=0)
-        if torch.isnan(result).any():
-            nan_frac = torch.isnan(result).float().mean().item()
-            logger.warning("NaN in prefill GDN layer %d (frac=%.4f), replacing with zeros",
-                           self.layer_idx, nan_frac)
-            result = torch.nan_to_num(result, nan=0.0)
-        return result
+        return output, state
 
-    def _do_decode(
-        self, hidden_states, attn_metadata, conv_state, temporal_state,
-        mixed_qkv_all, z_all, b_all, a_all,
-        conv1d_weight, A_log, dt_bias, norm, out_proj,
-        local_key_dim, local_val_dim, local_num_v, local_num_k, local_conv_dim,
-    ):
-        num_seqs = hidden_states.shape[0]
-        weight_2d = conv1d_weight.squeeze(1)
+    # ----- Single-step recurrent decode -----
+    def _single_step_decode(self, q, k, v, gate, beta, temporal_state):
+        if q.dim() == 4:
+            q = q.squeeze(1)
+            k = k.squeeze(1)
+            v = v.squeeze(1)
+            gate = gate.squeeze(1)
+            beta = beta.squeeze(1)
 
-        mixed_qkv = mixed_qkv_all.to(weight_2d.dtype).unsqueeze(-1)
-        mixed_qkv_conv = _causal_conv1d_update(
-            mixed_qkv, conv_state, weight_2d, bias=None, activation='silu')
-        mixed_qkv_conv = mixed_qkv_conv.squeeze(-1).unsqueeze(1)
+        B, H, D = q.shape
+        V = v.shape[-1]
 
-        q, k, v = torch.split(
-            mixed_qkv_conv,
-            [local_key_dim, local_key_dim, local_val_dim], dim=-1)
-        q = q.reshape(num_seqs, 1, local_num_k, self.head_k_dim)
-        k = k.reshape(num_seqs, 1, local_num_k, self.head_k_dim)
-        v = v.reshape(num_seqs, 1, local_num_v, self.head_v_dim)
+        q_f = F.normalize(q.float(), p=2, dim=-1)
+        k_f = F.normalize(k.float(), p=2, dim=-1)
+        v_f = v.float()
 
-        beta = b_all.sigmoid().unsqueeze(1)
-        _A_safe = A_log.float().clamp(-8.0, 4.0)
-        g = (-_A_safe.exp()
-             * F.softplus(a_all.float() + dt_bias).clamp(max=10.0)
-             ).unsqueeze(1)
+        if temporal_state is None:
+            temporal_state = torch.zeros(B, H, D, V,
+                                         dtype=torch.float32, device=q.device)
+        else:
+            temporal_state = temporal_state.float()
 
-        q = q.repeat_interleave(self.head_expand_ratio, dim=2)
-        k = k.repeat_interleave(self.head_expand_ratio, dim=2)
+        g = gate.float().clamp(-5.0, 0.0)
+        b = beta.float()
 
-        orig_dtype = q.dtype
-        _scale = self.head_k_dim ** -0.5
+        decay = torch.exp(g).unsqueeze(-1).unsqueeze(-1)
+        b_exp = b.unsqueeze(-1).unsqueeze(-1)
 
-        q_t = _l2norm(q.squeeze(1)).float() * _scale
-        k_t = _l2norm(k.squeeze(1)).float()
-        v_t = v.squeeze(1).float()
-        g_t = g.squeeze(1).float().clamp_(-20.0, 2.0).exp_()
-        bt = beta.squeeze(1).float()
+        kv = torch.einsum('bhd,bhv->bhdv', k_f, v_f)
+        temporal_state = decay * temporal_state + b_exp * kv
+        temporal_state = temporal_state.clamp(-100.0, 100.0)
 
-        temporal_state.mul_(g_t[:, :, None, None])
+        output = torch.einsum('bhd,bhdv->bhv', q_f, temporal_state)
+        output = output.clamp(-1e4, 1e4)
+        output = output.to(torch.float16).unsqueeze(1)
 
-        ts_flat = temporal_state.view(-1, self.head_k_dim, self.head_v_dim)
-        BH = ts_flat.shape[0]
-
-        kv_mem = _ix_bmm(
-            k_t.view(BH, 1, self.head_k_dim), ts_flat
-        ).view(num_seqs, local_num_v, self.head_v_dim)
-
-        delta = (v_t - kv_mem) * bt[:, :, None]
-
-        ts_flat.baddbmm_(
-            k_t.view(BH, self.head_k_dim, 1),
-            delta.view(BH, 1, self.head_v_dim),
-        )
-        temporal_state.clamp_(-65504.0, 65504.0)
-
-        core_out = _ix_bmm(
-            q_t.view(BH, 1, self.head_k_dim), ts_flat
-        ).view(num_seqs, local_num_v, self.head_v_dim).to(orig_dtype)
-
-        z = z_all.reshape(num_seqs, local_num_v, self.head_v_dim)
-        normed = norm(
-            core_out.reshape(-1, self.head_v_dim),
-            z.reshape(-1, self.head_v_dim))
-        normed = normed.reshape(num_seqs, -1)
-        out, _ = out_proj(normed)
-        return out
+        return output, temporal_state
diff --git a/qwen3_6_scripts/qwen3_5.py b/qwen3_6_scripts/qwen3_5.py
index f67d4a91..a07e0613 100644
--- a/qwen3_6_scripts/qwen3_5.py
+++ b/qwen3_6_scripts/qwen3_5.py
@@ -459,20 +459,24 @@ class GatedDeltaNet(nn.Module):
         self.norm = Qwen3_5RMSNormGated(self.head_v_dim,
                                         eps=text_cfg.rms_norm_eps)
 
-        # CoreX dispatch — our corex_gdn.py is deployed, init MUST succeed
+        # CoreX dispatch: try to create fused GDN operator from base image
         self._use_corex_gdn = False
         if _corex_gdn_available and _corex_gdn_module is not None:
-            self._corex_gdn_obj = _corex_gdn_module.CoreXGDN(
-                num_v_heads=self.num_v_heads // tp_size,
-                num_k_heads=self.num_k_heads // tp_size,
-                head_k_dim=self.head_k_dim,
-                head_v_dim=self.head_v_dim,
-                conv_kernel_size=self.conv_kernel_size,
-                layer_idx=layer_idx,
-            )
-            self._use_corex_gdn = True
-            if layer_idx == 0:
-                logger.info("GatedDeltaNet: CoreX fused GDN enabled")
+            try:
+                self._corex_gdn_obj = _corex_gdn_module.CoreXGDN(
+                    num_v_heads=self.num_v_heads // tp_size,
+                    num_k_heads=self.num_k_heads // tp_size,
+                    head_k_dim=self.head_k_dim,
+                    head_v_dim=self.head_v_dim,
+                    conv_kernel_size=self.conv_kernel_size,
+                    layer_idx=layer_idx,
+                )
+                self._use_corex_gdn = True
+                logger.info("GatedDeltaNet layer %d: CoreX fused GDN enabled", layer_idx)
+            except Exception as e:
+                logger.warning(
+                    "GatedDeltaNet layer %d: CoreX GDN init failed (%s), using PyTorch",
+                    layer_idx, e)
 
     def _conv1d_weight_loader(self, param: torch.Tensor,
                               loaded_weight: torch.Tensor) -> None:
@@ -498,16 +502,22 @@ class GatedDeltaNet(nn.Module):
         conv_state: torch.Tensor,          # (batch, local_conv_dim, kernel-1)  in-place
         temporal_state: torch.Tensor,      # (batch, local_v_heads, k_dim, v_dim)  in-place
     ) -> torch.Tensor:
-        # CoreX dispatch — NO FALLBACK. 0 score with fallback = same as crash.
+        # CoreX dispatch: try fused GDN kernel first (CCCL env_dispatch pattern)
         if self._use_corex_gdn:
-            return self._corex_gdn_obj.forward(
-                hidden_states, attn_metadata,
-                conv_state, temporal_state,
-                self.in_proj_qkv, self.in_proj_z,
-                self.in_proj_b, self.in_proj_a,
-                self.conv1d_weight, self.A_log, self.dt_bias,
-                self.norm, self.out_proj,
-            )
+            try:
+                return self._corex_gdn_obj.forward(
+                    hidden_states, attn_metadata,
+                    conv_state, temporal_state,
+                    self.in_proj_qkv, self.in_proj_z,
+                    self.in_proj_b, self.in_proj_a,
+                    self.conv1d_weight, self.A_log, self.dt_bias,
+                    self.norm, self.out_proj,
+                )
+            except Exception as e:
+                if self.layer_idx == 0:
+                    logger.warning(
+                        "CoreX GDN forward failed (%s), falling back", e)
+                self._use_corex_gdn = False  # permanent fallback
 
         # flash_qla SM70 DISABLED: produces inf on BI-V100 (abs mean=inf from real test)
         # xllm uses equivalent PyTorch chunked path (qwen3_gated_delta_net_base.cpp)
@@ -1069,17 +1079,20 @@ class Qwen3_5MoeSparseBlock(nn.Module):
         self.shared_expert_gate = ReplicatedLinear(
             hidden_size, 1, bias=False, quant_config=quant_config)
 
-        # CoreX dispatch — corex_moe.py is deployed, moe_forward MUST exist
+        # CoreX dispatch: try to use fused MoE kernels from base image
         self._use_corex_moe = False
         if _corex_moe_available and _corex_moe_module is not None:
-            self._corex_moe_forward = getattr(
-                _corex_moe_module, 'moe_forward', None)
-            if self._corex_moe_forward is not None:
-                self._use_corex_moe = True
-                if layer_idx == 0:
+            try:
+                # corex_moe module provides direct forward functions
+                self._corex_moe_forward = getattr(
+                    _corex_moe_module, 'moe_forward', None)
+                if self._corex_moe_forward is not None:
+                    self._use_corex_moe = True
                     logger.info("MoE: CoreX fused MoE forward available")
-            else:
-                raise RuntimeError("corex_moe module loaded but moe_forward missing")
+                else:
+                    logger.warning("MoE: corex_moe has no moe_forward, using PyTorch")
+            except Exception as e:
+                logger.warning("MoE: CoreX MoE init failed (%s), using PyTorch", e)
 
     def _pure_pytorch_experts(
         self,