fix: wire MoE topk via ixformer C++ bridge + disable broken flash_qla GDN
Two call chain breaks fixed: 1. MoE routing (2304 calls/token): BEFORE: torch.softmax + torch.topk (3 Python GPU ops, no ixformer) AFTER: ix_bridge.py → ix_moe_bridge.cpp → ixformer::infer::topk_softmax() Source: upstream_ref/xllm/xllm/core/kernels/ilu/fused_moe.cpp line 46 The C++ API exists in base image SDK (ixformer.h declares it), only the Python binding (ixformer.functions) was missing. 2. GDN prefill (4 layers, 99.98% NaN): BEFORE: flash_qla SM70 kernel → abs mean=inf → nan_to_num → zeros AFTER: skip flash_qla, use _pytorch_forward directly Source: upstream_ref/xllm qwen3_gated_delta_net_base.cpp uses identical PyTorch chunked logic (no flash_qla). Sub168 (working build) never deployed flash_qla either. Files: - ex_engine/csrc/ix_moe_bridge.cpp: torch C++ extension calling ixformer C++ API - ex_engine/python/ix_bridge.py: JIT-compile loader with PyTorch fallback - qwen3_5.py: import ix_bridge for MoE, disable flash_qla for GDN - patch_ops.sh: deploy ix_bridge .cpp + .py into vllm model dir
This commit is contained in:
EX Engine
@@ -178,23 +178,33 @@ if [ -n "$VLLM2" ]; then
|
||||
cp ./chat_utils.py "$VLLM2/entrypoints/chat_utils.py" 2>/dev/null || true
|
||||
fi
|
||||
|
||||
# Deploy EX Engine Python module into vllm importable path
|
||||
# Deploy EX Engine Python module + C++ bridge into vllm importable path
|
||||
EX_ENGINE_SRC="/workspace/ex_engine"
|
||||
if [ -d "$EX_ENGINE_SRC/python" ] && [ -d "$EX_ENGINE_SRC/build" ]; then
|
||||
if [ -d "$EX_ENGINE_SRC/python" ]; then
|
||||
# Deploy into vllm's model dir so qwen3_5.py can import it
|
||||
EX_DST="$VLLM/model_executor/models/ex_engine"
|
||||
mkdir -p "$EX_DST"
|
||||
cp "$EX_ENGINE_SRC/python/"*.py "$EX_DST/" 2>/dev/null || true
|
||||
# Copy built .so files
|
||||
cp "$EX_ENGINE_SRC/build/"*.so "$EX_DST/" 2>/dev/null || true
|
||||
echo "[patch_ops] EX Engine deployed: $(ls $EX_DST/*.so 2>/dev/null | wc -l) factors"
|
||||
mkdir -p "$EX_DST/python"
|
||||
mkdir -p "$EX_DST/csrc"
|
||||
cp "$EX_ENGINE_SRC/python/"*.py "$EX_DST/python/" 2>/dev/null || true
|
||||
# ix_moe_bridge.cpp needs to be next to the python module for JIT compile
|
||||
cp "$EX_ENGINE_SRC/csrc/ix_moe_bridge.cpp" "$EX_DST/csrc/" 2>/dev/null || true
|
||||
cp "$EX_ENGINE_SRC/csrc/ix_moe_bridge.cpp" "$EX_DST/python/" 2>/dev/null || true
|
||||
# Also make ex_engine importable from Python path
|
||||
touch "$EX_DST/__init__.py"
|
||||
touch "$EX_DST/python/__init__.py"
|
||||
# Copy built .so files if they exist
|
||||
if [ -d "$EX_ENGINE_SRC/build" ]; then
|
||||
cp "$EX_ENGINE_SRC/build/"*.so "$EX_DST/" 2>/dev/null || true
|
||||
fi
|
||||
echo "[patch_ops] EX Engine deployed to $EX_DST"
|
||||
ls -la "$EX_DST/csrc/" 2>/dev/null || true
|
||||
if [ -n "$VLLM2" ]; then
|
||||
EX_DST2="$VLLM2/model_executor/models/ex_engine"
|
||||
mkdir -p "$EX_DST2"
|
||||
cp "$EX_ENGINE_SRC/python/"*.py "$EX_DST2/" 2>/dev/null || true
|
||||
cp "$EX_ENGINE_SRC/build/"*.so "$EX_DST2/" 2>/dev/null || true
|
||||
mkdir -p "$EX_DST2/python" "$EX_DST2/csrc"
|
||||
cp -r "$EX_DST/"* "$EX_DST2/" 2>/dev/null || true
|
||||
fi
|
||||
else
|
||||
echo "[patch_ops] WARNING: EX Engine not built — MoE will use slow PyTorch fallback"
|
||||
echo "[patch_ops] WARNING: EX Engine not found — MoE uses slow PyTorch fallback"
|
||||
fi
|
||||
|
||||
echo "[patch_ops] DONE — EX Engine + SM70 GDN kernel + serving layer deployed"
|
||||
|
||||
@@ -71,8 +71,25 @@ except ImportError:
|
||||
|
||||
# corex_gdn/corex_moe: these are custom modules that teams package into their
|
||||
# Docker image. If present, they provide fused GDN/MoE kernels.
|
||||
_corex_gdn_module = None
|
||||
_corex_moe_module = None
|
||||
# ix_bridge: C++ bridge to ixformer::infer::topk_softmax (bypasses missing Python binding)
|
||||
_ix_bridge_module = None
|
||||
_ix_bridge_available = False
|
||||
try:
|
||||
from ex_engine.python.ix_bridge import topk_softmax as _ix_topk_softmax
|
||||
_ix_bridge_available = True
|
||||
logger.info("ix_bridge: ixformer C++ topk_softmax available")
|
||||
except ImportError:
|
||||
try:
|
||||
# Try deployed path inside vllm models dir
|
||||
import importlib, sys
|
||||
_ex_dir = os.path.join(os.path.dirname(__file__), "ex_engine")
|
||||
if os.path.isdir(_ex_dir) and _ex_dir not in sys.path:
|
||||
sys.path.insert(0, os.path.dirname(_ex_dir))
|
||||
from ex_engine.python.ix_bridge import topk_softmax as _ix_topk_softmax
|
||||
_ix_bridge_available = True
|
||||
logger.info("ix_bridge: ixformer C++ topk_softmax available (deployed path)")
|
||||
except ImportError:
|
||||
logger.info("ix_bridge: not available, MoE uses PyTorch topk")
|
||||
_corex_gdn_available = False
|
||||
_corex_moe_available = False
|
||||
|
||||
@@ -469,17 +486,13 @@ class GatedDeltaNet(nn.Module):
|
||||
"CoreX GDN forward failed (%s), falling back", e)
|
||||
self._use_corex_gdn = False # permanent fallback
|
||||
|
||||
# FlashQLA SM70 dispatch: fused CUDA kernel for prefill
|
||||
# Decode stays PyTorch (SM70 decode kernel needs different state layout)
|
||||
if _flash_qla_available and attn_metadata.num_prefill_tokens > 0:
|
||||
try:
|
||||
return self._flash_qla_prefill(
|
||||
hidden_states, attn_metadata, conv_state, temporal_state)
|
||||
except Exception as e:
|
||||
if self.layer_idx == 0:
|
||||
logger.warning(
|
||||
"FlashQLA SM70 prefill failed (%s), falling back to PyTorch", e)
|
||||
# Don't disable permanently — may work for different shapes
|
||||
# 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)
|
||||
# which works correctly in fp32. Keeping PyTorch path only.
|
||||
#
|
||||
# if _flash_qla_available and attn_metadata.num_prefill_tokens > 0:
|
||||
# try:
|
||||
# return self._flash_qla_prefill(...)
|
||||
|
||||
return self._pytorch_forward(
|
||||
hidden_states, attn_metadata, conv_state, temporal_state)
|
||||
@@ -1048,12 +1061,18 @@ class Qwen3_5MoeSparseBlock(nn.Module):
|
||||
Output is partial (pre-all-reduce), same contract as FusedMoE
|
||||
with reduce_results=False.
|
||||
"""
|
||||
# Routing: softmax → topk → renormalise
|
||||
routing_weights = _ix_softmax(router_logits.float(), dim=-1)
|
||||
topk_weights, topk_ids = torch.topk(
|
||||
routing_weights, self.top_k, dim=-1) # (T, top_k)
|
||||
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
||||
topk_weights = topk_weights.to(hidden_states.dtype)
|
||||
# Routing: fused topk+softmax via ixformer C++ bridge (if available)
|
||||
# Falls back to PyTorch softmax → topk → renormalize
|
||||
if _ix_bridge_available:
|
||||
topk_weights, topk_ids = _ix_topk_softmax(
|
||||
router_logits, self.top_k, renormalize=True)
|
||||
topk_weights = topk_weights.to(hidden_states.dtype)
|
||||
else:
|
||||
routing_weights = _ix_softmax(router_logits.float(), dim=-1)
|
||||
topk_weights, topk_ids = torch.topk(
|
||||
routing_weights, self.top_k, dim=-1) # (T, top_k)
|
||||
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
||||
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)
|
||||
|
||||
Reference in New Issue
Block a user