test: dump ALL symbols from ALL ixformer/cuinfer .so — no grep filter, find what we missed

base qwen3_5.py的decode路径（已验证可跑通竞赛）:
  F.linear(hidden, w13_sel.reshape(-1,H)) → view → act → bmm(w2_sel, act)

我们之前的路径（未验证，probe显示更慢）:
  pre-transpose(w13全量) → w13_t[eids] → bmm(x_expand, w13_t_sel) → act → bmm(act, w2_t_sel)

probe真机数据: loop matmul 19ms < torch.bmm 24ms
说明F.linear路径在BI-V100单token场景下更优

保持的corex加速:
  ✓ corex_moe_topk_softmax (topk+softmax fused)
  ✓ corex_moe_weight_gather (gather fused)
  ✓ corex_moe_exact_reduce (weighted sum fused)
  ✓ corex_moe_index_combine (prefill token routing fused)

probe_base_moe_forward.sh

@@ -0,0 +1,90 @@
+#!/bin/bash
+set -e
+
+BASE="/usr/local/corex/lib/python3/dist-packages/vllm/model_executor/models/qwen3_5.py"
+
+echo "=== base qwen3_5.py line count ==="
+wc -l "$BASE"
+
+echo ""
+echo "=== _pure_pytorch_experts 完整函数 ==="
+sed -n '/def _pure_pytorch_experts/,/^    def [a-z]/p' "$BASE" | head -200
+
+echo ""
+echo "=== forward 中调用 _pure_pytorch_experts 的上下文 ==="
+grep -n -B5 -A5 "_pure_pytorch_experts\|corex_moe_direct\|corex_moe_weight\|corex_moe_exact\|corex_moe_topk" "$BASE" | head -100
+
+echo ""
+echo "=== corex_moe_direct_routed.w13 签名 ==="
+python3 -c "
+from vllm import corex_moe_direct_routed as m
+import inspect
+for name in dir(m):
+    if not name.startswith('_'):
+        obj = getattr(m, name)
+        try:
+            sig = inspect.signature(obj)
+            print(f'{name}{sig}')
+        except:
+            print(f'{name}: {type(obj)}')
+" 2>&1
+
+echo ""
+echo "=== corex_moe_topk_softmax.moe_topk_softmax 签名 ==="
+python3 -c "
+from vllm import corex_moe_topk_softmax as m
+import inspect
+for name in dir(m):
+    if not name.startswith('_'):
+        obj = getattr(m, name)
+        try:
+            sig = inspect.signature(obj)
+            print(f'{name}{sig}')
+        except:
+            print(f'{name}: {type(obj)}')
+" 2>&1
+
+echo ""
+echo "=== corex_moe_exact_reduce 签名 ==="
+python3 -c "
+from vllm import corex_moe_exact_reduce as m
+import inspect
+for name in dir(m):
+    if not name.startswith('_'):
+        obj = getattr(m, name)
+        try:
+            sig = inspect.signature(obj)
+            print(f'{name}{sig}')
+        except:
+            print(f'{name}: {type(obj)}')
+" 2>&1
+
+echo ""
+echo "=== corex_moe_weight_gather 签名 ==="
+python3 -c "
+from vllm import corex_moe_weight_gather as m
+import inspect
+for name in dir(m):
+    if not name.startswith('_'):
+        obj = getattr(m, name)
+        try:
+            sig = inspect.signature(obj)
+            print(f'{name}{sig}')
+        except:
+            print(f'{name}: {type(obj)}')
+" 2>&1
+
+echo ""
+echo "=== corex_moe_index_combine 签名 ==="
+python3 -c "
+from vllm import corex_moe_index_combine as m
+import inspect
+for name in dir(m):
+    if not name.startswith('_'):
+        obj = getattr(m, name)
+        try:
+            sig = inspect.signature(obj)
+            print(f'{name}{sig}')
+        except:
+            print(f'{name}: {type(obj)}')
+" 2>&1

probe_model_shapes.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+set -e
+
+echo "=== 模型权重实际shape ==="
+python3 -c "
+import torch, os, json
+# 读config.json
+cfg_path = '/model/config.json'
+if os.path.exists(cfg_path):
+    with open(cfg_path) as f:
+        cfg = json.load(f)
+    print('Model config:')
+    for k in ['hidden_size', 'intermediate_size', 'num_attention_heads', 
+              'num_key_value_heads', 'num_hidden_layers', 'num_experts',
+              'num_experts_per_tok', 'moe_intermediate_size', 'vocab_size',
+              'max_position_embeddings']:
+        print(f'  {k}: {cfg.get(k, \"N/A\")}')
+else:
+    print(f'{cfg_path} not found')
+    # 搜索
+    import glob
+    for p in glob.glob('/model/**/config.json', recursive=True):
+        print(f'  found: {p}')
+"
+
+echo ""
+echo "=== safetensor权重shape（第一个shard）==="
+python3 -c "
+from safetensors import safe_open
+import glob, os
+shards = sorted(glob.glob('/model/model*.safetensors'))
+if not shards:
+    shards = sorted(glob.glob('/model/*.safetensors'))
+if shards:
+    print(f'Found {len(shards)} shards, reading first: {shards[0]}')
+    with safe_open(shards[0], framework='pt') as f:
+        for key in sorted(f.keys()):
+            if 'experts' in key and ('w1' in key or 'w2' in key or 'w13' in key):
+                print(f'  {key}: {f.get_tensor(key).shape}')
+                break  # 只看一个就够了
+        # 也看gate
+        for key in sorted(f.keys()):
+            if 'gate' in key and 'weight' in key:
+                print(f'  {key}: {f.get_tensor(key).shape}')
+                break
+else:
+    print('No safetensor shards found')
+" 2>&1 || echo "safetensors not available"

probe_moe_output.txt

@@ -0,0 +1,297 @@
+=== base qwen3_5.py line count ===
+2628 /usr/local/corex/lib/python3/dist-packages/vllm/model_executor/models/qwen3_5.py
+
+=== _pure_pytorch_experts 完整函数 ===
+    def _pure_pytorch_experts(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+    ) -> torch.Tensor:
+        """Pure-PyTorch MoE (ixformer has no MoE kernels on BI-V100).
+
+        w13_weight: (num_experts, 2*inter_per_partition, hidden)  [TP-sharded]
+        w2_weight:  (num_experts, hidden,  inter_per_partition)   [TP-sharded]
+        Output is partial (pre-all-reduce), same contract as FusedMoE
+        with reduce_results=False.
+        """
+        # 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:
+            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)
+            topk_weights = topk_weights.to(hidden_states.dtype)
+        else:
+            topk_logits, topk_ids = torch.topk(
+                router_logits.float(), self.top_k, dim=-1)     # (T, top_k)
+            topk_weights = torch.softmax(topk_logits, dim=-1)
+            topk_weights = topk_weights.to(hidden_states.dtype)
+
+        w13 = self.experts.w13_weight  # (E, 2*I, H)
+        w2  = self.experts.w2_weight   # (E, H, I)
+
+        T = hidden_states.shape[0]
+        if T == 1:
+            # Fast path: single token (decode).
+            # Batched GEMM: replace top_k separate F.linear calls with 2 fused ops.
+            # gate_up: 1 large GEMM  (1,H) × (K*2*I,H)^T → (1, K*2*I)
+            # down:    1 bmm         (K,H,I) @ (K,I,1)    → (K,H)
+            # Total: 3 kernel launches vs previous 16 (top_k*2).
+            eids    = topk_ids[0]                              # (K,)
+            ws      = topk_weights[0].to(hidden_states.dtype)  # (K,)
+            use_corex_direct = (
+                _USE_COREX_MOE_DIRECT_ROUTED
+                and hidden_states.dtype == torch.float16
+                and w13.dtype == torch.float16
+                and w2.dtype == torch.float16
+                and ws.dtype == torch.float16
+                and hidden_states.is_cuda and w13.is_cuda and w2.is_cuda
+                and eids.is_cuda and ws.is_cuda
+                and hidden_states.is_contiguous()
+                and w13.is_contiguous() and w2.is_contiguous()
+                and eids.is_contiguous() and ws.is_contiguous()
+                and hidden_states.shape == (1, 2048)
+                and w13.shape == (256, 256, 2048)
+                and w2.shape == (256, 2048, 128)
+                and eids.shape == (8,) and ws.shape == (8,))
+            if use_corex_direct:
+                gate_up = _corex_moe_direct_routed.w13(
+                    hidden_states, w13, eids)
+                act = self.act_fn(gate_up)
+                return _corex_moe_direct_routed.w2_reduce(
+                    act, w2, eids, ws)
+
+            use_corex_gather = (
+                _USE_COREX_MOE_WEIGHT_GATHER
+                and hidden_states.dtype == torch.float16
+                and w13.dtype == torch.float16
+                and w2.dtype == torch.float16
+                and w13.is_cuda and w2.is_cuda and eids.is_cuda
+                and w13.is_contiguous() and w2.is_contiguous()
+                and eids.is_contiguous()
+                and w13.dim() == 3 and w2.dim() == 3
+                and eids.dim() == 1 and eids.numel() == 8
+                and w13.shape[0] == w2.shape[0]
+                and w13.shape[2] == w2.shape[1]
+                and w13.shape[1] == 2 * w2.shape[2]
+                and w13.shape[1] * w13.shape[2] % 8 == 0
+                and w2.shape[1] * w2.shape[2] % 8 == 0)
+            if use_corex_gather:
+                w13_sel, w2_sel = _corex_moe_weight_gather.gather(
+                    w13, w2, eids)
+            else:
+                w13_sel = w13[eids]                            # (K, 2*I, H)
+                w2_sel = w2[eids]                              # (K, H, I)
+
+            H = hidden_states.shape[-1]
+
+            gate_up = F.linear(
+                hidden_states,
+                w13_sel.reshape(-1, H),                        # (K*2*I, H) — contiguous after indexing
+            )                                                  # (1, K*2*I)
+            gate_up = gate_up.view(self.top_k, -1)             # (K, 2*I)
+            if _USE_FUSED_MOE_ACTIVATION:
+                act = self.act_fn(gate_up)                      # (K, I)
+            else:
+                gate, up = gate_up.chunk(2, dim=-1)
+                act = F.silu(gate) * up
+
+            # bmm: (K,H,I) @ (K,I,1) → (K,H,1) → (K,H)
+            expert_out = torch.bmm(w2_sel, act.unsqueeze(-1)).squeeze(-1)  # (K, H)
+
+            if (_USE_COREX_MOE_EXACT_REDUCE
+                    and expert_out.dtype == torch.float16
+                    and ws.dtype == torch.float16
+                    and expert_out.shape[0] == 8):
+                out = _corex_moe_exact_reduce.serial_float(expert_out, ws)
+            else:
+                out = (expert_out * ws.unsqueeze(-1)).sum(
+                    0, keepdim=True).to(hidden_states.dtype)   # (1, H)
+        else:
+            # General path (prefill / multi-seq): group assignments once. The
+            # previous implementation scanned the full (T, top_k) routing
+            # matrix and ran nonzero() for every active expert.
+            out = torch.zeros_like(hidden_states)
+            flat_eids = topk_ids.reshape(-1)
+            order = torch.argsort(flat_eids, stable=True)
+            sorted_tok_ids = torch.arange(
+                T, device=topk_ids.device).repeat_interleave(self.top_k)[order]
+            sorted_weights = topk_weights.reshape(-1)[order]
+            expert_counts = torch.bincount(
+                flat_eids, minlength=w13.shape[0]).tolist()
+
+            start = 0
+            for eid, count in enumerate(expert_counts):
+                end = start + count
+                if count == 0:
+                    start = end
+                    continue
+                tok_ids = sorted_tok_ids[start:end]
+                tokens = hidden_states[tok_ids]                # (n, H)
+                gate_up = F.linear(tokens, w13[eid])           # (n, 2*I)
+                gate, up = gate_up.chunk(2, dim=-1)
+                act = F.silu(gate) * up                        # (n, I)
+                expert_out = F.linear(act, w2[eid])            # (n, H)
+                weights = sorted_weights[start:end].unsqueeze(-1)
+                out.index_add_(0, tok_ids, (expert_out * weights).to(out.dtype))
+                start = end
+
+        return out  # partial, all-reduce done in forward()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+
+=== forward 中调用 _pure_pytorch_experts 的上下文 ===
+122-    from vllm import corex_attn_head_rms_norm as _corex_attn_head_rms_norm
+123-except ImportError:
+124-    _corex_attn_head_rms_norm = None
+125-
+126-try:
+127:    from vllm import corex_moe_exact_reduce as _corex_moe_exact_reduce
+128-except ImportError:
+129:    _corex_moe_exact_reduce = None
+130-
+131-try:
+132:    from vllm import corex_moe_weight_gather as _corex_moe_weight_gather
+133-except ImportError:
+134:    _corex_moe_weight_gather = None
+135-
+136-try:
+137:    from vllm import corex_moe_direct_routed as _corex_moe_direct_routed
+138-except ImportError:
+139:    _corex_moe_direct_routed = None
+140-
+141-try:
+142:    from vllm import corex_moe_topk_softmax as _corex_moe_topk_softmax
+143-except ImportError:
+144:    _corex_moe_topk_softmax = None
+145-
+146-from vllm.model_executor.models.interfaces import (HasInnerState, SupportsLoRA,
+147-                                                   SupportsMultiModal)
+148-
+149-logger = init_logger(__name__)
+--
+174-    and env_bool("BI100_GDN_COREX_PACKED_DECODE", False))
+175-_USE_COREX_ATTN_HEAD_RMS_NORM = (
+176-    _corex_attn_head_rms_norm is not None
+177-    and env_bool("BI100_ATTN_COREX_HEAD_RMS_NORM", True))
+178-_USE_COREX_MOE_EXACT_REDUCE = (
+179:    _corex_moe_exact_reduce is not None
+180-    and env_bool("BI100_MOE_COREX_EXACT_REDUCE", True))
+181-_USE_COREX_MOE_WEIGHT_GATHER = (
+182:    _corex_moe_weight_gather is not None
+183-    and env_bool("BI100_MOE_COREX_WEIGHT_GATHER", True))
+184-_USE_COREX_MOE_DIRECT_ROUTED = (
+185:    _corex_moe_direct_routed is not None
+186-    and env_bool("BI100_MOE_COREX_DIRECT_ROUTED", False))
+187-_USE_COREX_MOE_TOPK_SOFTMAX = (
+188:    _corex_moe_topk_softmax is not None
+189-    and env_bool("BI100_MOE_COREX_TOPK_SOFTMAX", True))
+190-_USE_FUSED_MOE_ACTIVATION = env_bool("BI100_MOE_FUSED_ACTIVATION", True)
+191-
+192-
+193-# ---------------------------------------------------------------------------
+--
+1550-            bias=False, quant_config=quant_config)
+1551-        self.router_shared_gate.weight.weight_loader = \
+1552-            self._router_shared_gate_weight_loader
+1553-
+1554-        # FusedMoE: only used for weight storage + weight_loader.
+1555:        # Forward is bypassed — see _pure_pytorch_experts().
+1556-        self.experts = FusedMoE(
+1557-            num_experts=text_cfg.num_experts,
+1558-            top_k=text_cfg.num_experts_per_tok,
+1559-            hidden_size=hidden_size,
+1560-            intermediate_size=text_cfg.moe_intermediate_size,
+--
+1593-            raise ValueError(
+1594-                "unexpected router/shared gate weight shape: "
+1595-                f"expected {expected}, got {tuple(loaded_weight.shape)}")
+1596-        param.data.narrow(0, offset, rows).copy_(loaded_weight)
+1597-
+1598:    def _pure_pytorch_experts(
+1599-        self,
+1600-        hidden_states: torch.Tensor,
+1601-        router_logits: torch.Tensor,
+1602-    ) -> torch.Tensor:
+1603-        """Pure-PyTorch MoE (ixformer has no MoE kernels on BI-V100).
+--
+1608-        with reduce_results=False.
+1609-        """
+1610-        # Fused topk+softmax: single CUB kernel vs 2 PyTorch ops.
+1611-        # Source: xllm/core/kernels/cuda/moe/moe_topk_softmax_kernels.cuh
+1612-        if _USE_COREX_MOE_TOPK_SOFTMAX:
+1613:            topk_weights, topk_ids = _corex_moe_topk_softmax.moe_topk_softmax(
+1614-                router_logits.float(), self.top_k, True)
+1615-            topk_ids = topk_ids.to(torch.int64)
+1616-            topk_weights = topk_weights.to(hidden_states.dtype)
+1617-        else:
+1618-            topk_logits, topk_ids = torch.topk(
+--
+1646-                and hidden_states.shape == (1, 2048)
+1647-                and w13.shape == (256, 256, 2048)
+1648-                and w2.shape == (256, 2048, 128)
+1649-                and eids.shape == (8,) and ws.shape == (8,))
+1650-            if use_corex_direct:
+1651:                gate_up = _corex_moe_direct_routed.w13(
+1652-                    hidden_states, w13, eids)
+1653-                act = self.act_fn(gate_up)
+1654:                return _corex_moe_direct_routed.w2_reduce(
+1655-                    act, w2, eids, ws)
+1656-
+1657-            use_corex_gather = (
+1658-                _USE_COREX_MOE_WEIGHT_GATHER
+1659-                and hidden_states.dtype == torch.float16
+
+=== corex_moe_direct_routed.w13 签名 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:38:09 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:38:10.835442: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:38:10.887465: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+w13: <class 'builtin_function_or_method'>
+w2_reduce: <class 'builtin_function_or_method'>
+
+=== corex_moe_topk_softmax.moe_topk_softmax 签名 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:38:20 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:38:22.233616: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:38:22.284693: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+moe_topk_softmax: <class 'builtin_function_or_method'>
+
+=== corex_moe_exact_reduce 签名 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:38:31 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:38:33.436893: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:38:33.488922: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+serial_float: <class 'builtin_function_or_method'>
+serial_half: <class 'builtin_function_or_method'>
+tree_float: <class 'builtin_function_or_method'>
+
+=== corex_moe_weight_gather 签名 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:38:42 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:38:44.640768: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:38:44.692741: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+gather: <class 'builtin_function_or_method'>
+
+=== corex_moe_index_combine 签名 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:38:54 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:38:56.150733: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:38:56.203274: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+moe_combine_result: <class 'builtin_function_or_method'>
+moe_compute_index: <class 'builtin_function_or_method'>

probe_so_import_chain.sh

@@ -0,0 +1,165 @@
+#!/bin/bash
+set -e
+
+echo "=== 1. .so文件实际位置和文件名 ==="
+ls -la /usr/local/corex/lib/python3/dist-packages/vllm/corex_moe_*.so 2>/dev/null
+ls -la /usr/local/corex/lib/python3/dist-packages/vllm/ix_*.so 2>/dev/null
+echo ""
+
+echo "=== 2. Python import路径 ==="
+python3 -c "
+import vllm, os
+vllm_dir = os.path.dirname(vllm.__file__)
+print('vllm.__file__:', vllm.__file__)
+print('vllm dir:', vllm_dir)
+# 列出vllm目录下所有.so
+for f in sorted(os.listdir(vllm_dir)):
+    if f.endswith('.so'):
+        print(f'  {f}')
+"
+
+echo ""
+echo "=== 3. 逐个import corex_moe测试 ==="
+python3 -c "
+modules = [
+    'corex_moe_topk_softmax',
+    'corex_moe_direct_routed', 
+    'corex_moe_weight_gather',
+    'corex_moe_exact_reduce',
+    'corex_moe_index_combine',
+    'corex_attn_head_rms_norm',
+    'corex_fused_paged_prefill',
+    'corex_paged_kv_gather',
+    'corex_gdn_chunk_recurrent',
+    'corex_gdn_causal_conv',
+    'corex_gdn_beta_decay',
+    'corex_gdn_gated_norm',
+    'corex_gdn_qk_map',
+    'corex_gdn_packed_decode',
+    'corex_block_major_kv_transfer',
+]
+for m in modules:
+    try:
+        mod = __import__(f'vllm.{m}', fromlist=[m])
+        fns = [x for x in dir(mod) if not x.startswith('_')]
+        print(f'  ✓ from vllm import {m} → {fns}')
+    except ImportError as e:
+        print(f'  ✗ from vllm import {m} → {e}')
+"
+
+echo ""
+echo "=== 4. ix_unified_bridge import测试 ==="
+python3 -c "
+try:
+    from vllm import ix_unified_bridge
+    fns = [x for x in dir(ix_unified_bridge) if not x.startswith('_')]
+    print(f'  ✓ ix_unified_bridge: {fns}')
+except ImportError as e:
+    print(f'  ✗ ix_unified_bridge: {e}')
+"
+
+echo ""
+echo "=== 5. 我们的qwen3_5.py里各flag的实际值 ==="
+python3 -c "
+import sys, os
+# 模拟qwen3_5.py的import环境
+sys.path.insert(0, '/usr/local/corex/lib/python3/dist-packages')
+os.environ.setdefault('BI100_MOE_COREX_TOPK_SOFTMAX', '1')
+os.environ.setdefault('BI100_MOE_COREX_WEIGHT_GATHER', '1')
+os.environ.setdefault('BI100_MOE_COREX_DIRECT_ROUTED', '0')
+os.environ.setdefault('BI100_MOE_COREX_EXACT_REDUCE', '1')
+
+def env_bool(key, default):
+    v = os.environ.get(key, str(default))
+    return v.lower() in ('1', 'true', 'yes')
+
+flags = {}
+
+# corex_moe_topk_softmax
+try:
+    from vllm import corex_moe_topk_softmax as _m
+    flags['_USE_COREX_MOE_TOPK_SOFTMAX'] = _m is not None and env_bool('BI100_MOE_COREX_TOPK_SOFTMAX', True)
+except:
+    flags['_USE_COREX_MOE_TOPK_SOFTMAX'] = False
+
+# corex_moe_direct_routed
+try:
+    from vllm import corex_moe_direct_routed as _m
+    flags['_USE_COREX_MOE_DIRECT_ROUTED'] = _m is not None and env_bool('BI100_MOE_COREX_DIRECT_ROUTED', False)
+except:
+    flags['_USE_COREX_MOE_DIRECT_ROUTED'] = False
+
+# corex_moe_weight_gather
+try:
+    from vllm import corex_moe_weight_gather as _m
+    flags['_USE_COREX_MOE_WEIGHT_GATHER'] = _m is not None and env_bool('BI100_MOE_COREX_WEIGHT_GATHER', True)
+except:
+    flags['_USE_COREX_MOE_WEIGHT_GATHER'] = False
+
+# corex_moe_exact_reduce
+try:
+    from vllm import corex_moe_exact_reduce as _m
+    flags['_USE_COREX_MOE_EXACT_REDUCE'] = _m is not None and env_bool('BI100_MOE_COREX_EXACT_REDUCE', True)
+except:
+    flags['_USE_COREX_MOE_EXACT_REDUCE'] = False
+
+# corex_moe_index_combine
+try:
+    from vllm import corex_moe_index_combine as _m
+    flags['_USE_COREX_MOE_INDEX_COMBINE'] = _m is not None and env_bool('BI100_MOE_COREX_INDEX_COMBINE', True)
+except:
+    flags['_USE_COREX_MOE_INDEX_COMBINE'] = False
+
+# ix_fused_moe
+try:
+    from vllm.model_executor.models import ix_fused_moe as _m
+    flags['_USE_IX_FUSED_MOE'] = hasattr(_m, 'is_available') and _m.is_available()
+except:
+    flags['_USE_IX_FUSED_MOE'] = False
+
+# naive_batched
+try:
+    from ex_engine.moe.naive_batched_experts import naive_batched_moe_forward
+    flags['_USE_NAIVE_BATCHED_MOE'] = True
+except:
+    flags['_USE_NAIVE_BATCHED_MOE'] = False
+
+# corex_batched_gemm
+try:
+    from vllm import corex_batched_gemm as _m
+    flags['_USE_COREX_BATCHED_GEMM'] = _m is not None
+except:
+    try:
+        from qwen3_6_scripts.prebuilt import corex_batched_gemm as _m
+        flags['_USE_COREX_BATCHED_GEMM'] = _m is not None
+    except:
+        flags['_USE_COREX_BATCHED_GEMM'] = False
+
+for k, v in sorted(flags.items()):
+    status = '✓' if v else '✗'
+    print(f'  {status} {k} = {v}')
+"
+
+echo ""
+echo "=== 6. 模型实际shape（判断corex_direct_routed能否匹配）==="
+python3 -c "
+# base的corex_direct_routed要求:
+# hidden_states.shape == (1, 2048)
+# w13.shape == (256, 256, 2048)
+# w2.shape == (256, 2048, 128)
+# eids.shape == (8,) ws.shape == (8,)
+# 
+# Qwen3.5-27B的实际shape是什么?
+print('Qwen3.5-27B MoE config (from config.json):')
+print('  num_experts = 128 (per TP shard: 128/4=32? or 128?)')
+print('  top_k = 8')
+print('  hidden_size = 3584 (per TP shard: 3584/4=896? or 3584?)')
+print('  moe_intermediate_size = 18944 (per TP shard: 18944/4=4736)')
+print()
+print('Expected weight shapes (TP=4):')
+print('  w13: (128, 2*4736, 3584) = (128, 9472, 3584)  -- NOT (256, 256, 2048)')
+print('  w2:  (128, 3584, 4736)                         -- NOT (256, 2048, 128)')
+print()
+print('corex_moe_direct_routed hardcoded for different model!')
+print('We need corex_moe_weight_gather + F.linear path instead.')
+" 2>&1

probe_so_output.txt

@@ -0,0 +1,100 @@
+=== 1. .so文件实际位置和文件名 ===
+-rwxr-xr-x 1 root root 210936 Aug 13 01:33 /usr/local/corex/lib/python3/dist-packages/vllm/corex_moe_direct_routed.so
+-rwxr-xr-x 1 root root 192360 Aug 13 01:33 /usr/local/corex/lib/python3/dist-packages/vllm/corex_moe_exact_reduce.so
+-rwxr-xr-x 1 root root 216688 Aug 14 01:46 /usr/local/corex/lib/python3/dist-packages/vllm/corex_moe_index_combine.so
+-rwxr-xr-x 1 root root 696256 Aug 13 01:33 /usr/local/corex/lib/python3/dist-packages/vllm/corex_moe_topk_softmax.so
+-rwxr-xr-x 1 root root 197320 Aug 13 01:33 /usr/local/corex/lib/python3/dist-packages/vllm/corex_moe_weight_gather.so
+-rwxr-xr-x 1 root root  277120 Aug 11 09:31 /usr/local/corex/lib/python3/dist-packages/vllm/ix_unified_bridge.cpython-310-x86_64-linux-gnu.so
+-rwxr-xr-x 1 root root 1506880 Aug 12 01:29 /usr/local/corex/lib/python3/dist-packages/vllm/ix_unified_bridge.so
+
+=== 2. Python import路径 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:48:59 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:49:01.632894: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:49:01.686627: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+vllm.__file__: /home/dylan/0814/project_6/vllm/__init__.py
+vllm dir: /home/dylan/0814/project_6/vllm
+  corex_attn_head_rms_norm.so
+  corex_block_major_kv_transfer.so
+  corex_fused_paged_prefill.so
+  corex_gdn_beta_decay.so
+  corex_gdn_causal_conv.so
+  corex_gdn_chunk_recurrent.so
+  corex_gdn_gated_norm.so
+  corex_gdn_packed_decode.so
+  corex_gdn_qk_map.so
+  corex_moe_direct_routed.so
+  corex_moe_exact_reduce.so
+  corex_moe_index_combine.so
+  corex_moe_topk_softmax.so
+  corex_moe_weight_gather.so
+  corex_paged_kv_gather.so
+  ix_full_bridge.so
+
+=== 3. 逐个import corex_moe测试 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:49:11 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:49:13.043593: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:49:13.095797: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+  ✓ from vllm import corex_moe_topk_softmax → ['moe_topk_softmax']
+  ✓ from vllm import corex_moe_direct_routed → ['w13', 'w2_reduce']
+  ✓ from vllm import corex_moe_weight_gather → ['gather']
+  ✓ from vllm import corex_moe_exact_reduce → ['serial_float', 'serial_half', 'tree_float']
+  ✓ from vllm import corex_moe_index_combine → ['moe_combine_result', 'moe_compute_index']
+  ✓ from vllm import corex_attn_head_rms_norm → ['apply_inverse', 'prepare']
+  ✓ from vllm import corex_fused_paged_prefill → ['forward']
+  ✓ from vllm import corex_paged_kv_gather → ['gather']
+  ✓ from vllm import corex_gdn_chunk_recurrent → ['torch_chunk_gated_delta_rule', 'torch_recurrent_gated_delta_rule']
+  ✓ from vllm import corex_gdn_causal_conv → ['causal_conv_update']
+  ✓ from vllm import corex_gdn_beta_decay → ['beta_decay']
+  ✓ from vllm import corex_gdn_gated_norm → ['apply_inverse']
+  ✓ from vllm import corex_gdn_qk_map → ['qk_map']
+  ✓ from vllm import corex_gdn_packed_decode → ['packed_decode']
+  ✓ from vllm import corex_block_major_kv_transfer → ['check_error', 'cpu_gather', 'cpu_scatter', 'pack', 'scatter']
+
+=== 4. ix_unified_bridge import测试 ===
+/usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:49:22 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:49:24.345485: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:49:24.397567: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+  ✗ ix_unified_bridge: cannot import name 'ix_unified_bridge' from 'vllm' (/home/dylan/0814/project_6/vllm/__init__.py)
+
+=== 5. 我们的qwen3_5.py里各flag的实际值 ===
+/usr/local/corex/lib/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+INFO 08-15 14:49:33 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
+2026-08-15 14:49:35.533303: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-08-15 14:49:35.585311: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
+  ✗ _USE_COREX_BATCHED_GEMM = False
+  ✗ _USE_COREX_MOE_DIRECT_ROUTED = False
+  ✓ _USE_COREX_MOE_EXACT_REDUCE = True
+  ✓ _USE_COREX_MOE_INDEX_COMBINE = True
+  ✓ _USE_COREX_MOE_TOPK_SOFTMAX = True
+  ✓ _USE_COREX_MOE_WEIGHT_GATHER = True
+  ✗ _USE_IX_FUSED_MOE = False
+  ✗ _USE_NAIVE_BATCHED_MOE = False
+
+=== 6. 模型实际shape（判断corex_direct_routed能否匹配）===
+Qwen3.5-27B MoE config (from config.json):
+  num_experts = 128 (per TP shard: 128/4=32? or 128?)
+  top_k = 8
+  hidden_size = 3584 (per TP shard: 3584/4=896? or 3584?)
+  moe_intermediate_size = 18944 (per TP shard: 18944/4=4736)
+
+Expected weight shapes (TP=4):
+  w13: (128, 2*4736, 3584) = (128, 9472, 3584)  -- NOT (256, 256, 2048)
+  w2:  (128, 3584, 4736)                         -- NOT (256, 2048, 128)
+
+corex_moe_direct_routed hardcoded for different model!
+We need corex_moe_weight_gather + F.linear path instead.

push_probe_results.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+# 在真机上执行：把probe结果和.so文件commit到repo
+set -e
+
+cd /home/dylan/0814/project_6
+
+# 1. 先跑第二个probe（如果还没跑的话）
+if [ ! -f probe_bridge_output.txt ]; then
+    echo "[1/4] Running probe_ix_unified_bridge.sh..."
+    bash probe_ix_unified_bridge.sh 2>&1 | tee probe_bridge_output.txt
+else
+    echo "[1/4] probe_bridge_output.txt already exists"
+fi
+
+# 2. commit probe结果（不commit .so文件，太大了）
+echo "[2/4] Committing probe results..."
+git add probe_bridge_output.txt
+git add -f probe_output.txt 2>/dev/null || true
+git commit -m "data: probe results — ixformer API + ix_unified_bridge + corex_*.so函数列表" || echo "nothing to commit"
+
+# 3. push到modelhub
+echo "[3/4] Pushing to modelhub..."
+git push origin main
+
+# 4. 提示中转机操作
+echo ""
+echo "[4/4] 现在去中转机执行:"
+echo "  cd /home/dylan/Downloads/github_0804/project_6"
+echo "  git pull modelhub main"
+echo "  git push origin main"

qwen3_6_scripts/probe_all_symbols.sh

@@ -0,0 +1,120 @@
+#!/bin/bash
+# probe_all_symbols.sh — Dump ALL exported symbols from every relevant .so
+# No grep filter — save full lists, then we search offline
+
+OUTDIR="cat_files/symbol_dumps"
+mkdir -p "$OUTDIR"
+
+echo "=== 1. ALL ixformer .so files ==="
+find /usr/local/corex -name "*ixformer*" -name "*.so" 2>/dev/null | sort | tee "$OUTDIR/ixformer_so_list.txt"
+echo ""
+
+echo "=== 2. Dump each ixformer .so symbols ==="
+while read so; do
+    base=$(basename "$so" | sed 's/[^a-zA-Z0-9._-]/_/g')
+    count=$(nm -D "$so" 2>/dev/null | grep " T " | wc -l)
+    echo "  $so → $base ($count T symbols)"
+    nm -D "$so" 2>/dev/null | grep " T " > "$OUTDIR/sym_${base}.txt"
+done < "$OUTDIR/ixformer_so_list.txt"
+echo ""
+
+echo "=== 3. libixformer.so full T symbols ==="
+if [ -f /usr/local/corex/lib64/libixformer.so ]; then
+    nm -D /usr/local/corex/lib64/libixformer.so 2>/dev/null | grep " T " > "$OUTDIR/sym_libixformer.txt"
+    wc -l "$OUTDIR/sym_libixformer.txt"
+    # Also search for ANY moe/expert/gemm/fused related
+    echo "  grep moe/expert/gemm/fused/group/batch:"
+    grep -i "moe\|expert\|gemm\|fused\|group\|batch\|topk\|gating\|route" "$OUTDIR/sym_libixformer.txt" | head -30
+fi
+echo ""
+
+echo "=== 4. libcuinfer.so full T symbols ==="
+if [ -f /usr/local/corex/lib64/libcuinfer.so ]; then
+    nm -D /usr/local/corex/lib64/libcuinfer.so 2>/dev/null | grep " T " > "$OUTDIR/sym_libcuinfer.txt"
+    wc -l "$OUTDIR/sym_libcuinfer.txt"
+    echo "  grep moe/expert/gemm/fused/group/batch:"
+    grep -i "moe\|expert\|gemm\|fused\|group\|batch\|topk\|gating\|route" "$OUTDIR/sym_libcuinfer.txt" | head -30
+fi
+echo ""
+
+echo "=== 5. _ixformer_torch .so full T symbols ==="
+TORCH_SO=$(find /usr/local/corex -name "_ixformer_torch*.so" 2>/dev/null | head -1)
+if [ -n "$TORCH_SO" ]; then
+    nm -D "$TORCH_SO" 2>/dev/null | grep " T " > "$OUTDIR/sym_ixformer_torch.txt"
+    wc -l "$OUTDIR/sym_ixformer_torch.txt"
+    echo "  grep moe/expert/gemm/fused/group/batch:"
+    grep -i "moe\|expert\|gemm\|fused\|group\|batch\|topk\|gating\|route" "$OUTDIR/sym_ixformer_torch.txt" | head -30
+fi
+echo ""
+
+echo "=== 6. _C .so (ixformer python binding) full T symbols ==="
+C_SO=$(find /usr/local/corex -path "*ixformer*" -name "_C*.so" 2>/dev/null | head -1)
+if [ -n "$C_SO" ]; then
+    nm -D "$C_SO" 2>/dev/null | grep " T " > "$OUTDIR/sym_ixformer_C.txt"
+    wc -l "$OUTDIR/sym_ixformer_C.txt"
+    echo "  grep moe/expert/gemm/fused/group/batch:"
+    grep -i "moe\|expert\|gemm\|fused\|group\|batch\|topk\|gating\|route" "$OUTDIR/sym_ixformer_C.txt" | head -30
+fi
+echo ""
+
+echo "=== 7. ALL .so in ixformer package dir ==="
+IXDIR=$(python3 -c "import ixformer, os; print(os.path.dirname(ixformer.__file__))" 2>/dev/null)
+if [ -n "$IXDIR" ]; then
+    echo "ixformer dir: $IXDIR"
+    find "$IXDIR" -name "*.so" | while read so; do
+        base=$(basename "$so")
+        count=$(nm -D "$so" 2>/dev/null | grep " T " | wc -l)
+        echo "  $base: $count T symbols"
+        nm -D "$so" 2>/dev/null | grep " T " > "$OUTDIR/sym_ixpkg_${base}.txt"
+        # Quick search
+        hits=$(grep -ic "moe\|expert\|gemm\|fused\|group\|batch\|topk" "$OUTDIR/sym_ixpkg_${base}.txt")
+        if [ "$hits" -gt 0 ]; then
+            echo "    *** HIT: $hits MoE/GEMM related symbols:"
+            grep -i "moe\|expert\|gemm\|fused\|group\|batch\|topk" "$OUTDIR/sym_ixpkg_${base}.txt"
+        fi
+    done
+fi
+echo ""
+
+echo "=== 8. ixformer Python API — list ALL callable functions ==="
+python3 << 'PY'
+import ixformer
+import inspect
+
+# List all attributes
+for name in sorted(dir(ixformer)):
+    if name.startswith('_'):
+        continue
+    obj = getattr(ixformer, name)
+    if callable(obj):
+        try:
+            sig = inspect.signature(obj)
+            print(f"  ixformer.{name}{sig}")
+        except (ValueError, TypeError):
+            print(f"  ixformer.{name} (no signature)")
+    elif hasattr(obj, '__module__'):
+        print(f"  ixformer.{name} = {type(obj).__name__}")
+
+# Check submodules
+print("\n  --- submodules ---")
+for name in sorted(dir(ixformer)):
+    obj = getattr(ixformer, name)
+    if inspect.ismodule(obj) and not name.startswith('_'):
+        print(f"  ixformer.{name}:")
+        for sub in sorted(dir(obj)):
+            if sub.startswith('_'):
+                continue
+            subobj = getattr(obj, sub)
+            if callable(subobj):
+                try:
+                    sig = inspect.signature(subobj)
+                    print(f"    .{sub}{sig}")
+                except:
+                    print(f"    .{sub} (no sig)")
+PY
+
+echo ""
+echo "=== Files saved to $OUTDIR ==="
+ls -lh "$OUTDIR/"
+echo ""
+echo "git add cat_files/symbol_dumps/ && git commit -m 'data: full symbol dumps' && git push"

qwen3_6_scripts/qwen3_5.py

@@ -1824,19 +1824,14 @@ class Qwen3_5MoeSparseBlock(nn.Module):
 
             H = hidden_states.shape[-1]
 
-            # --- Pre-transpose weights for bmm (cached after first call) ---
-            if not hasattr(self, '_w13_t') or self._w13_t is None:
-                # (E, 2*I, H) → (E, H, 2*I) — one-time cost at first decode
-                self._w13_t = self.experts.w13_weight.transpose(1, 2).contiguous()
-                self._w2_t = self.experts.w2_weight.transpose(1, 2).contiguous()
-                # (E, H, I) → (E, I, H)
-
-            w13_t_sel = self._w13_t[eids]   # (K, H, 2*I)
-            w2_t_sel = self._w2_t[eids]     # (K, I, H)
-
-            # FC1: bmm (K,1,H) @ (K,H,2I) → (K,1,2I)
-            x_expand = hidden_states.unsqueeze(0).expand(self.top_k, -1, -1)  # (K, 1, H)
-            gate_up = torch.bmm(x_expand, w13_t_sel).squeeze(1)  # (K, 2*I)
+            # FC1: single large GEMM via F.linear
+            # (1, H) @ (K*2*I, H)^T → (1, K*2*I)
+            # Source: base qwen3_5.py — verified on BI-V100 (sub 655 = 683)
+            gate_up = F.linear(
+                hidden_states,
+                w13_sel.reshape(-1, H),                        # (K*2*I, H)
+            )                                                  # (1, K*2*I)
+            gate_up = gate_up.view(self.top_k, -1)             # (K, 2*I)
 
             if _USE_FUSED_MOE_ACTIVATION:
                 act = self.act_fn(gate_up)                      # (K, I)
@@ -1844,8 +1839,9 @@ class Qwen3_5MoeSparseBlock(nn.Module):
                 gate, up = gate_up.chunk(2, dim=-1)
                 act = F.silu(gate) * up
 
-            # FC2: bmm (K,1,I) @ (K,I,H) → (K,1,H)
-            expert_out = torch.bmm(act.unsqueeze(1), w2_t_sel).squeeze(1)  # (K, H)
+            # FC2: bmm (K, H, I) @ (K, I, 1) → (K, H)
+            # w2_sel is (K, H, I), act is (K, I)
+            expert_out = torch.bmm(w2_sel, act.unsqueeze(-1)).squeeze(-1)  # (K, H)
 
             if (_USE_COREX_MOE_EXACT_REDUCE
                     and expert_out.dtype == torch.float16