a54dbda3bb
- ix_bridge.py: auto-discover ixformer .so files, pass as extra_ldflags - ix_moe_bridge.cpp: fix mangled header from bad sed, add #include <optional> - verify_single_gpu.py: also pass extra_ldflags during JIT compile The undefined symbol _ZN8ixformer5infer12silu_and_mulERN2at6TensorES3_ lives in libixformer.so — need to explicitly link it.
460 lines
16 KiB
Python
460 lines
16 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
verify_single_gpu.py — 单卡 BI-V100 验证 ix_full_bridge + MoE dispatch chain
|
|
|
|
用法: python3 verify_single_gpu.py
|
|
要求: 在真机 Docker 里运行，单卡即可
|
|
|
|
测试链条:
|
|
Step 0: JIT 编译 ix_full_bridge.cpp → .so
|
|
Step 1: ixformer::infer::topk_softmax
|
|
Step 2: ixformer::infer::moe_compute_token_index_api (gen_idx)
|
|
Step 3: ixformer::infer::moe_expand_input
|
|
Step 4: ixformer::infer::moe_w16a16_group_gemm (w13)
|
|
Step 5: ixformer::infer::silu_and_mul
|
|
Step 6: ixformer::infer::moe_w16a16_group_gemm (w2)
|
|
Step 7: ixformer::infer::moe_output_reduce_sum (combine)
|
|
Step 8: fused_moe_forward (全链路一次调用)
|
|
Step 9: paged_attention
|
|
Step 10: flash_attn_prefill
|
|
Step 11: rms_norm
|
|
"""
|
|
|
|
import os
|
|
import sys
|
|
import time
|
|
import traceback
|
|
|
|
# ============================================================================
|
|
# Step 0: JIT compile ix_full_bridge.cpp
|
|
# ============================================================================
|
|
def step0_compile_bridge():
|
|
print("=" * 60)
|
|
print("STEP 0: JIT compile ix_full_bridge.cpp")
|
|
print("=" * 60)
|
|
|
|
# Find the source
|
|
here = os.path.dirname(os.path.abspath(__file__))
|
|
candidates = [
|
|
os.path.join(here, "ex_engine", "csrc", "ix_full_bridge.cpp"),
|
|
"/workspace/ex_engine/csrc/ix_full_bridge.cpp",
|
|
]
|
|
cpp_path = None
|
|
for c in candidates:
|
|
if os.path.exists(c):
|
|
cpp_path = c
|
|
break
|
|
if cpp_path is None:
|
|
print(" ✗ ix_full_bridge.cpp NOT FOUND")
|
|
print(f" Searched: {candidates}")
|
|
return None
|
|
|
|
print(f" Source: {cpp_path}")
|
|
|
|
from torch.utils.cpp_extension import load
|
|
import glob
|
|
|
|
# Find ixformer .so to link against
|
|
extra_ldflags = []
|
|
try:
|
|
import ixformer
|
|
ixf_dir = os.path.dirname(ixformer.__file__)
|
|
for so in glob.glob(os.path.join(ixf_dir, "*.so")):
|
|
if "cpython" not in so:
|
|
extra_ldflags.append(so)
|
|
for so in glob.glob(os.path.join(ixf_dir, "_ixformer_torch*.so")):
|
|
extra_ldflags.append(so)
|
|
except ImportError:
|
|
pass
|
|
corex_lib = "/usr/local/corex/lib64"
|
|
if os.path.isdir(corex_lib):
|
|
for lib in ["libixattn.so", "libixformer.so"]:
|
|
p = os.path.join(corex_lib, lib)
|
|
if os.path.exists(p) and p not in extra_ldflags:
|
|
extra_ldflags.append(p)
|
|
extra_ldflags.append(f"-Wl,-rpath,{corex_lib}")
|
|
try:
|
|
import ixformer
|
|
extra_ldflags.append(f"-Wl,-rpath,{os.path.dirname(ixformer.__file__)}")
|
|
except ImportError:
|
|
pass
|
|
|
|
print(f" Link libs: {[os.path.basename(x) for x in extra_ldflags if not x.startswith('-')]}")
|
|
|
|
t0 = time.time()
|
|
try:
|
|
bridge = load(
|
|
name="ix_full_bridge",
|
|
sources=[cpp_path],
|
|
extra_cflags=["-O2", "-std=c++17"],
|
|
extra_ldflags=extra_ldflags,
|
|
verbose=True,
|
|
)
|
|
dt = time.time() - t0
|
|
fns = [x for x in dir(bridge) if not x.startswith("_")]
|
|
print(f" ✓ Compiled in {dt:.1f}s")
|
|
print(f" Functions: {fns}")
|
|
return bridge
|
|
except Exception as e:
|
|
dt = time.time() - t0
|
|
print(f" ✗ Compile FAILED after {dt:.1f}s")
|
|
print(f" Error: {e}")
|
|
traceback.print_exc()
|
|
return None
|
|
|
|
|
|
# ============================================================================
|
|
# Step 1-7: MoE dispatch chain (individual steps)
|
|
# ============================================================================
|
|
def step1_topk_softmax(bridge):
|
|
import torch
|
|
print("\nSTEP 1: topk_softmax")
|
|
gating = torch.randn(4, 64, dtype=torch.float32, device="cuda") # 4 tokens, 64 experts
|
|
topk = 8
|
|
try:
|
|
weights, ids = bridge.topk_softmax(gating, topk, True)
|
|
print(f" ✓ weights: {weights.shape} {weights.dtype}, ids: {ids.shape} {ids.dtype}")
|
|
print(f" weights sum per token: {weights.sum(dim=-1).tolist()}")
|
|
print(f" ids range: [{ids.min().item()}, {ids.max().item()}]")
|
|
assert weights.shape == (4, 8)
|
|
assert ids.shape == (4, 8)
|
|
assert ids.max().item() < 64
|
|
return weights, ids
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return None, None
|
|
|
|
|
|
def step2_gen_idx(bridge, ids):
|
|
import torch
|
|
print("\nSTEP 2: moe_gen_idx")
|
|
try:
|
|
flat_ids = ids.view(-1) # (32,)
|
|
result = bridge.moe_gen_idx(flat_ids, 64)
|
|
print(f" ✓ Returns {len(result)} tensors:")
|
|
names = ["src_dst", "dst_src", "expert_sizes", "cumsum"]
|
|
for i, (name, t) in enumerate(zip(names, result)):
|
|
print(f" [{i}] {name}: {t.shape} {t.dtype}")
|
|
return result
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return None
|
|
|
|
|
|
def step3_expand_input(bridge, idx, topk=8):
|
|
import torch
|
|
print("\nSTEP 3: moe_expand_input")
|
|
hidden = torch.randn(4, 1536, dtype=torch.float16, device="cuda") # 4 tokens, hidden=1536
|
|
try:
|
|
expanded = bridge.moe_expand_input(hidden, idx[0], idx[1], topk)
|
|
print(f" ✓ expanded: {expanded.shape} {expanded.dtype}")
|
|
assert expanded.shape[0] == 4 * topk
|
|
return expanded, hidden
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return None, None
|
|
|
|
|
|
def step4_group_gemm_w13(bridge, expanded, idx):
|
|
import torch
|
|
print("\nSTEP 4: group_gemm (w13)")
|
|
# w13: (64 experts, 2*intermediate, hidden) — simulate small
|
|
# Real: (64, 4608, 1536) but we use smaller for test
|
|
E, inter2, H = 64, 256, 1536
|
|
w13 = torch.randn(E, inter2, H, dtype=torch.float16, device="cuda")
|
|
try:
|
|
gemm1 = bridge.group_gemm(expanded, w13, idx[2], inter2)
|
|
print(f" ✓ gemm1: {gemm1.shape} {gemm1.dtype}")
|
|
return gemm1, w13
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return None, None
|
|
|
|
|
|
def step5_silu_and_mul(bridge, gemm1):
|
|
import torch
|
|
print("\nSTEP 5: silu_and_mul")
|
|
try:
|
|
act = bridge.silu_and_mul(gemm1)
|
|
print(f" ✓ act: {act.shape} {act.dtype}")
|
|
assert act.shape[-1] == gemm1.shape[-1] // 2
|
|
return act
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return None
|
|
|
|
|
|
def step6_group_gemm_w2(bridge, act, idx):
|
|
import torch
|
|
print("\nSTEP 6: group_gemm (w2)")
|
|
E, H, I = 64, 1536, act.shape[-1]
|
|
w2 = torch.randn(E, H, I, dtype=torch.float16, device="cuda")
|
|
try:
|
|
gemm2 = bridge.group_gemm(act, w2, idx[2], H)
|
|
print(f" ✓ gemm2: {gemm2.shape} {gemm2.dtype}")
|
|
return gemm2
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return None
|
|
|
|
|
|
def step7_combine_result(bridge, gemm2, weights):
|
|
import torch
|
|
print("\nSTEP 7: moe_combine_result")
|
|
try:
|
|
result = bridge.moe_combine_result(gemm2, weights)
|
|
print(f" ✓ result: {result.shape} {result.dtype}")
|
|
print(f" NaN: {result.isnan().any().item()}, Inf: {result.isinf().any().item()}")
|
|
return result
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return None
|
|
|
|
|
|
# ============================================================================
|
|
# Step 8: Full fused_moe_forward
|
|
# ============================================================================
|
|
def step8_fused_moe(bridge):
|
|
import torch
|
|
print("\n" + "=" * 60)
|
|
print("STEP 8: fused_moe_forward (FULL PIPELINE)")
|
|
print("=" * 60)
|
|
num_tokens = 4
|
|
hidden_size = 1536
|
|
num_experts = 64
|
|
inter_size = 128 # small for test
|
|
topk = 8
|
|
|
|
hidden = torch.randn(num_tokens, hidden_size, dtype=torch.float16, device="cuda")
|
|
router = torch.randn(num_tokens, num_experts, dtype=torch.float16, device="cuda")
|
|
w13 = torch.randn(num_experts, inter_size * 2, hidden_size, dtype=torch.float16, device="cuda")
|
|
w2 = torch.randn(num_experts, hidden_size, inter_size, dtype=torch.float16, device="cuda")
|
|
|
|
try:
|
|
result = bridge.fused_moe_forward(hidden, router, w13, w2, topk, num_experts, True)
|
|
print(f" ✓ result: {result.shape} {result.dtype}")
|
|
print(f" NaN: {result.isnan().any().item()}, Inf: {result.isinf().any().item()}")
|
|
print(f" abs_mean: {result.abs().mean().item():.4f}")
|
|
return True
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return False
|
|
|
|
|
|
# ============================================================================
|
|
# Step 9: paged_attention
|
|
# ============================================================================
|
|
def step9_paged_attention(bridge):
|
|
import torch
|
|
print("\nSTEP 9: paged_attention")
|
|
B, Hq, Hkv, D = 1, 4, 1, 128
|
|
block_size = 16
|
|
max_blocks = 4
|
|
seq_len = 48 # fits in 3 blocks
|
|
|
|
query = torch.randn(B, Hq, D, dtype=torch.float16, device="cuda")
|
|
# KV cache: (num_blocks, Hkv, block_size, D)
|
|
total_blocks = max_blocks
|
|
key_cache = torch.randn(total_blocks, Hkv, block_size, D, dtype=torch.float16, device="cuda")
|
|
value_cache = torch.randn(total_blocks, Hkv, block_size, D, dtype=torch.float16, device="cuda")
|
|
block_tables = torch.arange(max_blocks, dtype=torch.int32, device="cuda").unsqueeze(0)
|
|
seq_lens = torch.tensor([seq_len], dtype=torch.int32, device="cuda")
|
|
output = torch.empty(B, Hq, D, dtype=torch.float16, device="cuda")
|
|
|
|
try:
|
|
bridge.paged_attention(
|
|
output, query, key_cache, value_cache,
|
|
Hkv, D ** -0.5,
|
|
block_tables, seq_lens, block_size, seq_len, None)
|
|
print(f" ✓ output: {output.shape} {output.dtype}")
|
|
print(f" NaN: {output.isnan().any().item()}, abs_mean: {output.abs().mean().item():.4f}")
|
|
return True
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return False
|
|
|
|
|
|
# ============================================================================
|
|
# Step 10: flash_attn_prefill
|
|
# ============================================================================
|
|
def step10_flash_attn(bridge):
|
|
import torch
|
|
print("\nSTEP 10: flash_attn_prefill")
|
|
Hq, Hkv, D = 4, 1, 128
|
|
seq_len = 32
|
|
|
|
query = torch.randn(seq_len, Hq, D, dtype=torch.float16, device="cuda")
|
|
key = torch.randn(seq_len, Hkv, D, dtype=torch.float16, device="cuda")
|
|
value = torch.randn(seq_len, Hkv, D, dtype=torch.float16, device="cuda")
|
|
output = torch.empty_like(query)
|
|
block_tables = torch.empty(0, dtype=torch.int32, device="cuda")
|
|
cu_q = torch.tensor([0, seq_len], dtype=torch.int32, device="cuda")
|
|
cu_k = torch.tensor([0, seq_len], dtype=torch.int32, device="cuda")
|
|
|
|
try:
|
|
bridge.flash_attn_prefill(
|
|
query, key, value, output, block_tables,
|
|
cu_q, cu_k, seq_len, seq_len, D ** -0.5, True, -1, -1)
|
|
print(f" ✓ output: {output.shape} {output.dtype}")
|
|
print(f" NaN: {output.isnan().any().item()}, abs_mean: {output.abs().mean().item():.4f}")
|
|
return True
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return False
|
|
|
|
|
|
# ============================================================================
|
|
# Step 11: rms_norm
|
|
# ============================================================================
|
|
def step11_rms_norm(bridge):
|
|
import torch
|
|
print("\nSTEP 11: rms_norm")
|
|
hidden = torch.randn(4, 1536, dtype=torch.float16, device="cuda")
|
|
weight = torch.ones(1536, dtype=torch.float16, device="cuda")
|
|
output = torch.empty_like(hidden)
|
|
try:
|
|
bridge.rms_norm(output, hidden, weight, 1e-6)
|
|
print(f" ✓ output: {output.shape}")
|
|
print(f" NaN: {output.isnan().any().item()}, abs_mean: {output.abs().mean().item():.4f}")
|
|
return True
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return False
|
|
|
|
|
|
# ============================================================================
|
|
# Step 12: corex_moe.py Python module (tests tiered dispatch)
|
|
# ============================================================================
|
|
def step12_corex_moe_python():
|
|
import torch
|
|
print("\n" + "=" * 60)
|
|
print("STEP 12: corex_moe.py Python tiered dispatch")
|
|
print("=" * 60)
|
|
|
|
# Add project root to path
|
|
here = os.path.dirname(os.path.abspath(__file__))
|
|
sys.path.insert(0, here)
|
|
|
|
try:
|
|
from ex_engine.python.corex_moe import moe_forward, topk_softmax
|
|
print(" ✓ corex_moe imported")
|
|
except Exception as e:
|
|
print(f" ✗ Import failed: {e}")
|
|
return False
|
|
|
|
num_tokens = 4
|
|
hidden_size = 256 # small for test
|
|
num_experts = 8 # small
|
|
inter_size = 64
|
|
topk = 2
|
|
|
|
hidden = torch.randn(num_tokens, hidden_size, dtype=torch.float16, device="cuda")
|
|
gate = torch.randn(num_tokens, num_experts, dtype=torch.float16, device="cuda")
|
|
w13 = torch.randn(num_experts, inter_size * 2, hidden_size, dtype=torch.float16, device="cuda")
|
|
w2 = torch.randn(num_experts, hidden_size, inter_size, dtype=torch.float16, device="cuda")
|
|
|
|
try:
|
|
result = moe_forward(hidden, gate, w13, w2, topk=topk,
|
|
renormalize=True, num_experts=num_experts)
|
|
print(f" ✓ moe_forward: {result.shape} {result.dtype}")
|
|
print(f" NaN: {result.isnan().any().item()}, abs_mean: {result.abs().mean().item():.4f}")
|
|
return True
|
|
except Exception as e:
|
|
print(f" ✗ FAILED: {e}")
|
|
traceback.print_exc()
|
|
return False
|
|
|
|
|
|
# ============================================================================
|
|
# Main
|
|
# ============================================================================
|
|
def main():
|
|
import torch
|
|
print("=" * 60)
|
|
print(" BI-V100 Single GPU Verification")
|
|
print(f" CUDA available: {torch.cuda.is_available()}")
|
|
if torch.cuda.is_available():
|
|
print(f" Device: {torch.cuda.get_device_name(0)}")
|
|
print(f" Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
|
|
print("=" * 60)
|
|
|
|
results = {}
|
|
|
|
# Step 0: Compile
|
|
bridge = step0_compile_bridge()
|
|
results["compile"] = bridge is not None
|
|
|
|
if bridge is not None:
|
|
# Steps 1-7: Individual MoE steps
|
|
weights, ids = step1_topk_softmax(bridge)
|
|
results["topk_softmax"] = weights is not None
|
|
|
|
if ids is not None:
|
|
idx = step2_gen_idx(bridge, ids)
|
|
results["gen_idx"] = idx is not None
|
|
|
|
if idx is not None:
|
|
expanded, hidden = step3_expand_input(bridge, idx)
|
|
results["expand_input"] = expanded is not None
|
|
|
|
if expanded is not None:
|
|
gemm1, w13 = step4_group_gemm_w13(bridge, expanded, idx)
|
|
results["group_gemm_w13"] = gemm1 is not None
|
|
|
|
if gemm1 is not None:
|
|
act = step5_silu_and_mul(bridge, gemm1)
|
|
results["silu_and_mul"] = act is not None
|
|
|
|
if act is not None:
|
|
gemm2 = step6_group_gemm_w2(bridge, act, idx)
|
|
results["group_gemm_w2"] = gemm2 is not None
|
|
|
|
if gemm2 is not None:
|
|
result = step7_combine_result(bridge, gemm2, weights)
|
|
results["combine_result"] = result is not None
|
|
|
|
# Step 8: Full pipeline
|
|
results["fused_moe"] = step8_fused_moe(bridge)
|
|
|
|
# Step 9-11: Other kernels
|
|
results["paged_attention"] = step9_paged_attention(bridge)
|
|
results["flash_attn"] = step10_flash_attn(bridge)
|
|
results["rms_norm"] = step11_rms_norm(bridge)
|
|
|
|
# Step 12: Python module test (works even without bridge)
|
|
results["corex_moe_python"] = step12_corex_moe_python()
|
|
|
|
# Summary
|
|
print("\n" + "=" * 60)
|
|
print(" SUMMARY")
|
|
print("=" * 60)
|
|
for name, ok in results.items():
|
|
status = "✓ PASS" if ok else "✗ FAIL"
|
|
print(f" {status} {name}")
|
|
|
|
passed = sum(1 for v in results.values() if v)
|
|
total = len(results)
|
|
print(f"\n {passed}/{total} passed")
|
|
|
|
if results.get("fused_moe"):
|
|
print("\n >>> MoE FULL C++ PIPELINE WORKS — comp 168 parity achieved <<<")
|
|
elif results.get("corex_moe_python"):
|
|
print("\n >>> MoE Python fallback works — C++ bridge needs debugging <<<")
|
|
|
|
return 0 if passed == total else 1
|
|
|
|
|
|
if __name__ == "__main__":
|
|
sys.exit(main())
|