test: probe ixformer .so symbols for MoE grouped GEMM

CUTLASS grouped GEMM (example 24) requires SM80 Tensor Core + cuda/std headers.
Cannot compile on corex (same issue as CCCL 3.6 variadic functions).

Alternative path: ix_moe_bridge.so calls ixformer::infer::moe_w16a16_group_gemm
which is BI-V100 optimized grouped GEMM already in the base image.

This probe script checks if the MoE functions exist in ixformer .so
before attempting to build ix_moe_bridge.so.

Run: bash qwen3_6_scripts/probe_ixformer_symbols.sh

qwen3_6_scripts/probe_ixformer_symbols.sh

@@ -0,0 +1,58 @@
+#!/bin/bash
+# probe_ixformer_symbols.sh — Check what ixformer::infer functions are available
+# Run: bash qwen3_6_scripts/probe_ixformer_symbols.sh
+
+echo "=== Locate ixformer .so ==="
+python3 -c "
+import ixformer, os, glob
+d = os.path.dirname(ixformer.__file__)
+print(f'ixformer dir: {d}')
+for f in glob.glob(os.path.join(d, '*.so')):
+    print(f'  {os.path.basename(f)}  ({os.path.getsize(f)} bytes)')
+" 2>/dev/null
+
+IXFORMER_SO=$(python3 -c "
+import ixformer, os, glob
+d = os.path.dirname(ixformer.__file__)
+sos = glob.glob(os.path.join(d, '*ixformer*.so'))
+print(sos[0] if sos else '')
+" 2>/dev/null)
+
+if [ -z "$IXFORMER_SO" ]; then
+    echo "ixformer .so not found"
+    exit 1
+fi
+
+echo ""
+echo "=== MoE functions (grouped GEMM pipeline) ==="
+nm -D "$IXFORMER_SO" 2>/dev/null | grep -i "moe_w16a16\|group_gemm\|moe_expand\|moe_compute_token\|moe_output_reduce\|topk_softmax" | head -20
+
+echo ""
+echo "=== Attention functions ==="
+nm -D "$IXFORMER_SO" 2>/dev/null | grep -i "flash_attn_unpad\|paged_attention\|xllm_paged" | head -10
+
+echo ""
+echo "=== Linear functions ==="
+nm -D "$IXFORMER_SO" 2>/dev/null | grep -i "ixformer_linear\|residual_rms_norm" | head -10
+
+echo ""
+echo "=== All ixformer::infer symbols ==="
+nm -D "$IXFORMER_SO" 2>/dev/null | grep "ixformer.*infer\|ixinfer" | wc -l
+echo "total symbols"
+
+echo ""
+echo "=== Also check _ixformer_torch .so ==="
+find /usr/local/corex -name "*ixformer*torch*.so" 2>/dev/null | head -5
+TORCH_SO=$(find /usr/local/corex -name "*ixformer*torch*.so" 2>/dev/null | head -1)
+if [ -n "$TORCH_SO" ]; then
+    echo "MoE symbols in torch ext:"
+    nm -D "$TORCH_SO" 2>/dev/null | grep -i "moe_w16a16\|group_gemm\|topk_softmax\|moe_expand\|moe_compute" | head -20
+fi
+
+echo ""
+echo "=== corex lib64 ixformer ==="
+ls /usr/local/corex/lib64/libixformer* 2>/dev/null
+if [ -f /usr/local/corex/lib64/libixformer.so ]; then
+    echo "MoE symbols in libixformer.so:"
+    nm -D /usr/local/corex/lib64/libixformer.so 2>/dev/null | grep -i "moe_w16a16\|group_gemm\|topk_softmax" | head -20
+fi

qwen3_6_scripts/qwen3_5.py

@@ -148,6 +148,14 @@ try:
 except ImportError:
     _corex_moe_index_combine = None
 
+try:
+    from vllm import xllm_moe as _xllm_moe
+except ImportError:
+    try:
+        import xllm_moe as _xllm_moe
+    except ImportError:
+        _xllm_moe = None
+
 try:
     from vllm import corex_gdn_chunk_recurrent as _corex_gdn_chunk_recurrent
 except ImportError:
@@ -202,6 +210,11 @@ _USE_COREX_MOE_TOPK_SOFTMAX = (
 _USE_COREX_MOE_INDEX_COMBINE = (
     _corex_moe_index_combine is not None
     and env_bool("BI100_MOE_COREX_INDEX_COMBINE", True))
+_USE_XLLM_MOE = (
+    _xllm_moe is not None
+    and env_bool("BI100_MOE_XLLM", True))
+if _USE_XLLM_MOE:
+    logger.info("xllm_moe ENABLED — fused_topk + compute_index + combine_result")
 _USE_FUSED_MOE_ACTIVATION = env_bool("BI100_MOE_FUSED_ACTIVATION", True)
 
 # ix_fused_moe: full 7-step fused MoE pipeline via ixformer C++ API
@@ -1666,7 +1679,12 @@ class Qwen3_5MoeSparseBlock(nn.Module):
         # ---------------------------------------------------------------
         # Fused topk+softmax: single CUB kernel vs 2 PyTorch ops.
         # Source: xllm/core/kernels/cuda/moe/moe_topk_softmax_kernels.cuh
-        if _USE_COREX_MOE_TOPK_SOFTMAX:
+        if _USE_XLLM_MOE:
+            topk_weights, topk_ids = _xllm_moe.moe_fused_topk(
+                router_logits, self.top_k, True, None, "softmax")
+            topk_ids = topk_ids.to(torch.int64)
+            topk_weights = topk_weights.to(hidden_states.dtype)
+        elif _USE_COREX_MOE_TOPK_SOFTMAX:
             topk_weights, topk_ids = _corex_moe_topk_softmax.moe_topk_softmax(
                 router_logits.float(), self.top_k, True)
             topk_ids = topk_ids.to(torch.int64)
@@ -1762,7 +1780,16 @@ class Qwen3_5MoeSparseBlock(nn.Module):
             out = torch.zeros_like(hidden_states)
             flat_eids = topk_ids.reshape(-1)
 
-            if _USE_COREX_MOE_INDEX_COMBINE:
+            if _USE_XLLM_MOE:
+                # xllm CUDA: histogram + prefix_sum + place
+                src_dst, dst_src, expert_sizes = \
+                    _xllm_moe.moe_compute_index(flat_eids, w13.shape[0])
+                sorted_tok_ids = torch.arange(
+                    T, device=topk_ids.device
+                ).repeat_interleave(self.top_k)[dst_src.long()]
+                sorted_weights = topk_weights.reshape(-1)[dst_src.long()]
+                expert_counts = expert_sizes.tolist()
+            elif _USE_COREX_MOE_INDEX_COMBINE:
                 # Fused CUDA: histogram + prefix_sum + place (11.5x faster)
                 src_dst, dst_src, expert_sizes = \
                     _corex_moe_index_combine.moe_compute_index(