test: verify_build.sh for real hardware validation

- import ix_unified bridge到qwen3_5.py
- MoE prefill: ix_bridge.moe_group_gemm替代per-expert for-loop
- 保留fallback: ix_bridge失败自动回退到PyTorch for-loop
- 新feature flag: BI100_MOE_IX_BRIDGE (default=True when bridge available)

qwen3_6_scripts/qwen3_5.py

@@ -143,7 +143,15 @@ from vllm.model_executor.models.interfaces import (HasInnerState, SupportsLoRA,
 
 logger = init_logger(__name__)
 
-_bi100_model_trace("qwen3_5 runtime imports complete")
+# --- ix_unified: bridge to ixformer::infer C++ APIs -------------------------
+try:
+    from ix_unified import ix as _ix_bridge
+    _HAS_IX_BRIDGE = (_ix_bridge._bridge is not None)
+except ImportError:
+    _ix_bridge = None
+    _HAS_IX_BRIDGE = False
+
+_bi100_model_trace(f"qwen3_5 runtime imports complete (ix_bridge={_HAS_IX_BRIDGE})")
 
 _ALLOW_GDN_NAN_ZERO = env_bool("BI100_GDN_ALLOW_NAN_ZERO", False)
 _GDN_FINITE_CHECK = (env_bool("BI100_GDN_FINITE_CHECK", False)
@@ -180,6 +188,7 @@ _USE_COREX_MOE_DIRECT_ROUTED = (
     _corex_moe_direct_routed is not None
     and env_bool("BI100_MOE_COREX_DIRECT_ROUTED", False))
 _USE_FUSED_MOE_ACTIVATION = env_bool("BI100_MOE_FUSED_ACTIVATION", True)
+_USE_IX_BRIDGE_MOE = (_HAS_IX_BRIDGE and env_bool("BI100_MOE_IX_BRIDGE", True))
 
 
 # ---------------------------------------------------------------------------
@@ -1688,18 +1697,32 @@ class Qwen3_5MoeSparseBlock(nn.Module):
                 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.
+            # General path (prefill / multi-seq)
             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()
+            expert_counts_t = torch.bincount(
+                flat_eids, minlength=w13.shape[0])
 
+            if _USE_IX_BRIDGE_MOE:
+                # ix_bridge path: batched group_gemm instead of per-expert loop
+                try:
+                    sorted_hidden = hidden_states[sorted_tok_ids]
+                    gate_up = _ix_bridge.moe_group_gemm(
+                        sorted_hidden, w13, expert_counts_t)
+                    act = self.act_fn(gate_up)
+                    down = _ix_bridge.moe_group_gemm(
+                        act, w2, expert_counts_t)
+                    down_weighted = down * sorted_weights.unsqueeze(-1)
+                    out.index_add_(0, sorted_tok_ids, down_weighted.to(out.dtype))
+                    return out
+                except Exception as e:
+                    logger.warning("ix_bridge MoE failed (%s), falling back", e)
+
+            expert_counts = expert_counts_t.tolist()
             start = 0
             for eid, count in enumerate(expert_counts):
                 end = start + count

verify_build.sh

@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+# verify_build.sh — 在BI-V100真机上验证完整build链
+set -uo pipefail
+cd "$(dirname "$0")"
+
+echo "=== 1. patch_ops.sh syntax ==="
+bash -n qwen3_6_scripts/patch_ops.sh && echo "✓ OK" || echo "✗ FAIL"
+
+echo ""
+echo "=== 2. build ix_unified_bridge.so ==="
+bash ex_engine/build_unified_bridge.sh 2>&1 | tail -10
+
+echo ""
+echo "=== 3. verify bridge load ==="
+python3 << 'PY'
+import importlib.util, glob
+so = glob.glob("ex_engine/build/ix_unified_bridge*.so")
+if not so:
+    print("✗ .so not built"); exit(1)
+spec = importlib.util.spec_from_file_location("ix_unified_bridge", so[0])
+mod = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(mod)
+funcs = [x for x in dir(mod) if not x.startswith('_')]
+print(f"✓ {len(funcs)} functions: {funcs}")
+PY
+
+echo ""
+echo "=== 4. ix_unified dispatch smoke test ==="
+python3 << 'PY'
+import sys, os
+sys.path.insert(0, "ex_engine/build")
+sys.path.insert(0, "ex_engine/python")
+os.environ["IX_BRIDGE_PATH"] = "ex_engine/build"
+from ix_unified import ix
+print(f"bridge={ix._bridge is not None}, ixformer={ix._ixf is not None}")
+
+import torch
+# silu_and_mul
+x = torch.randn(2, 512, device="cuda", dtype=torch.float16)
+out = ix.silu_and_mul(x)
+print(f"✓ silu_and_mul: {x.shape} → {out.shape}")
+
+# rms_norm
+inp = torch.randn(4, 2048, device="cuda", dtype=torch.float16)
+outp = torch.empty_like(inp)
+w = torch.ones(2048, device="cuda", dtype=torch.float16)
+ix.rms_norm(outp, inp, w, 1e-6)
+print(f"✓ rms_norm: {inp.shape}")
+
+# moe_topk_softmax
+gate = torch.randn(8, 256, device="cuda", dtype=torch.float16)
+weights, indices = ix.moe_topk_softmax(gate, 8, True)
+print(f"✓ moe_topk_softmax: {gate.shape} → weights={weights.shape} indices={indices.shape}")
+
+print("ALL SMOKE TESTS PASSED")
+PY
+
+echo ""
+echo "=== 5. prebuilt .so still load ==="
+python3 << 'PY'
+import importlib.util, os
+so_dir = "qwen3_6_scripts/prebuilt/corex-3.2.3-ivcore10"
+for name in ["corex_moe_direct_routed", "corex_gdn_packed_decode", "corex_gdn_causal_conv"]:
+    so = os.path.join(so_dir, f"{name}.so")
+    spec = importlib.util.spec_from_file_location(name, so)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    funcs = [x for x in dir(mod) if not x.startswith('_')]
+    print(f"✓ {name}: {funcs}")
+PY
+
+echo ""
+echo "=== 6. py_compile all ==="
+cd qwen3_6_scripts
+find . -path './wheels' -prune -o -name '*.py' -print0 | xargs -0 python3 -m py_compile 2>&1 && echo "✓ all OK" || echo "✗ errors"
+
+echo ""
+echo "=== DONE ==="