fix(deltanet): systematic NaN elimination via CCCL overflow_cast pattern
Sub509 docker logs show 99.98-100% NaN rate in ALL GatedDeltaNet layers. nan_to_num(nan=0.0) replaces them with zeros — entire DeltaNet layers produce zero output, crippling model quality. This is root cause of: - d03_tool_call FAIL (model too impaired to output <tool_call> XML) - d01 content[0] (no content output, only 1085 reasoning tokens) - d04 content[0] (same: reasoning but no content) - 11x slower than opponent (model generates excessive tokens) Five-layer fix based on CCCL overflow_cast_t pattern: 1. Pre-cumsum clamp: g.clamp(-0.5, 0.5) before cumsum (was: no pre-clamp) - Limits cumsum growth to ±32 for chunk_size=64 - Post-cumsum clamp tightened from ±20 to ±12 2. A_log clamp tightened: [-20,20] → [-5,5] - exp(5) ≈ 148 vs exp(20) ≈ 4.9e8 - Prevents extreme decay rates that feed into g 3. Forward substitution per-row clamp: ±1e4 - _forward_sub_lower was the primary NaN amplifier - Each x[i] = rhs[i] + A[i,:i]@x[:i] now clamped 4. Cross-chunk state clamp: ±1e4 - last_state *= g_last_exp can blow up across many chunks - Both prefill and decode paths protected 5. Decode temporal_state in-place clamp: ±1e4 - ts_flat.clamp_() after baddbmm_ state update Also adds SUB509_DEEP_DIAGNOSIS.md with full root cause analysis.
This commit is contained in:
Claude
141
SUB509_DEEP_DIAGNOSIS.md
Normal file
141
SUB509_DEEP_DIAGNOSIS.md
Normal file
@@ -0,0 +1,141 @@
|
||||
# Sub509 深度诊断 — 基于CCCL源码阅读的系统级分析
|
||||
|
||||
## 一、Sub509 vs Sub168 关键数据对比
|
||||
|
||||
| 测试 | 对手Sub168 | 我们Sub509 | 差距分析 |
|
||||
|------|-----------|-----------|---------|
|
||||
| d01_basic_nostream | 8.49s, content[11] tok=139 | 95.85s, content[0] reasoning[1102] tok=1085 | 11x慢; 我们产了1085个token全是reasoning |
|
||||
| d02_stream_usage | 2.75s, chunks=53 | 1.84s, chunks=9 | 我们居然更快(但只产了9个chunks vs 53) |
|
||||
| d03_tool_call | 2.12s, tool=get_weather | **49.04s, tools=0 finish=stop** | **致命**: 模型不输出<tool_call> XML |
|
||||
| d04_reasoning | 17.78s, content[181] reasoning[1011] | 128.74s, content[0] reasoning[1447] | 7x慢; 我们有reasoning但没有content |
|
||||
|
||||
## 二、三大根因(按严重程度排序)
|
||||
|
||||
### 根因1: GatedDeltaNet每层产NaN → 模型"智力"丧失
|
||||
|
||||
docker日志证据:
|
||||
```
|
||||
WARNING qwen3_5.py:445] NaN in prefill GatedDeltaNet layer 0 (frac=0.9998)
|
||||
WARNING qwen3_5.py:445] NaN in prefill GatedDeltaNet layer 1 (frac=0.9997)
|
||||
WARNING qwen3_5.py:445] NaN in prefill GatedDeltaNet layer 2 (frac=1.0000)
|
||||
WARNING qwen3_5.py:445] NaN in prefill GatedDeltaNet layer 4 (frac=1.0000)
|
||||
```
|
||||
|
||||
**99.98%-100% NaN率**。`nan_to_num(result, nan=0.0)` 将这些NaN替换为零，等于整个DeltaNet层输出全是零。
|
||||
这是一种"活着但脑死亡"的状态——前向传播不报错，但模型失去了DeltaNet层的能力。
|
||||
|
||||
**NaN来源追踪**:
|
||||
1. `_torch_chunk_gated_delta_rule` 中 `g.cumsum(dim=-1)` → 累积值可能极大
|
||||
2. `g.clamp(-20,20)` 后 `g.exp()` → 最大 ~5e8，但这些值进入矩阵乘法后仍可能溢出
|
||||
3. `decay_mask = (g_diff).tril().exp()` → 即使单个exp不溢出，大矩阵乘法的累加也可能溢出
|
||||
4. `_forward_sub_lower` 中的前向替代: `x[i] = rhs[i] + A[i,:i] @ x[:i]`，如果A中有大值，误差逐行放大
|
||||
|
||||
**对手为什么没有这个问题**: 对手可能用的是不同的模型架构(不含DeltaNet)，或者在NVIDIA GPU上float32精度够高不会溢出。
|
||||
|
||||
### 根因2: FusedMoE完全fallback → 性能灾难
|
||||
|
||||
```
|
||||
ERROR _custom_ops.py:58] module 'ixformer.functions' has no attribute 'vllm_moe_topk_softmax'
|
||||
WARNING qwen3_5.py:913] FusedMoE native kernel failed, falling back to pure PyTorch experts permanently.
|
||||
```
|
||||
|
||||
BI-V100的ixformer没有MoE kernel，所有MoE层都用纯PyTorch:
|
||||
- 256个expert × top_k=8 → 最多256次F.linear调用(prefill)
|
||||
- 每次decode也需要top_k=8次expert forward
|
||||
- 对比native kernel的1次fused launch，这是数量级的差距
|
||||
|
||||
### 根因3: computility-run.yaml vs 实际参数不一致
|
||||
|
||||
yaml写的: `--max-model-len 256000 --max-num-seqs 2 --gpu-memory-utilization 0.95`
|
||||
docker日志: `max_seq_len=100000, max_num_seqs=1, gpu_memory_utilization=0.9`
|
||||
|
||||
**可能原因**: 部署时还在用旧的配置。需要确认yaml是否真的被用于部署。
|
||||
|
||||
## 三、d03_tool_call为什么FAIL
|
||||
|
||||
d03日志: `tools=0 finish=stop reasoning[0] (tool_choice=auto) (49.04s)`
|
||||
|
||||
**reasoning[0]说明enable_thinking=False确实生效了**。但模型仍然不输出`<tool_call>` XML。
|
||||
|
||||
analysis:
|
||||
1. enable_thinking=False → 模型不产生`<think>...</think>`块 ✓
|
||||
2. 但模型的输出内容不包含`<tool_call><function=get_weather>...` 格式
|
||||
3. tool parser `Qwen3CoderToolParser` 在输出中找不到 `<function=` → tools_called=False
|
||||
4. 49.04s意味着模型在漫长生成纯文本回答(可能是口头描述天气而不是调用tool)
|
||||
|
||||
**核心问题: GatedDeltaNet的NaN导致模型质量太差,不能正确follow tool_call格式**
|
||||
|
||||
这不是serving_chat.py的问题。serving_chat.py和protocol.py中的tool_call thinking禁用逻辑是正确的。问题在模型本身。
|
||||
|
||||
## 四、对手Sub168分析
|
||||
|
||||
对手最终得分60194.6:
|
||||
- functional: ~48/52 PASS
|
||||
- case_truncation: score=1.0
|
||||
- replay_tencent: score=60194 (94/881成功, tps avg 11.86)
|
||||
- opencompass: 0.0 (server也崩了)
|
||||
|
||||
**对手server在replay后期也崩溃了**(704个connection refused)。
|
||||
**对手的replay也只有94/881成功(10.7%)**。
|
||||
|
||||
但对手赢在:
|
||||
1. functional高通过率 → 基础分
|
||||
2. case_truncation通过 → 引擎稳定
|
||||
3. replay中94个成功请求 × tps → 得到分数
|
||||
|
||||
## 五、修复路径(按投入产出比排序)
|
||||
|
||||
### 修复1: NaN问题 — 强制float32精度 + 更激进的clamp
|
||||
|
||||
当前: `g.clamp(-20, 20)` 不够。cumsum后再clamp太晚了。
|
||||
需要: 在cumsum之前就对g的原始值做clamp。
|
||||
|
||||
在 `_torch_chunk_gated_delta_rule`:
|
||||
```python
|
||||
# 现在: g = g.cumsum(dim=-1).clamp(-20, 20)
|
||||
# 改为: g = g.clamp(-5, 5).cumsum(dim=-1).clamp(-15, 15)
|
||||
```
|
||||
|
||||
在 GatedDeltaNet.forward 的 prefill path:
|
||||
```python
|
||||
# 现在: _A_safe = self.A_log.float().clamp(-20.0, 20.0)
|
||||
# 改为更窄: _A_safe = self.A_log.float().clamp(-10.0, 10.0)
|
||||
```
|
||||
|
||||
### 修复2: MoE性能 — 尝试真正使用native kernel
|
||||
|
||||
docker日志说 `vllm_moe_topk_softmax` 不存在，但 _custom_ops.py 里应该有PyTorch fallback。
|
||||
问题是 `_hw_policy.moe_native_align` 和 `_hw_policy.moe_native_invoke` 也是False。
|
||||
如果这两个真的不存在，那PyTorch fallback就是唯一选择。
|
||||
|
||||
**性能改进**: 在 `_pure_pytorch_experts` 的 prefill path 中:
|
||||
- 现在: for-loop over experts, 每个一次F.linear
|
||||
- 改为: 按expert batch size排序，大batch的expert合并成一个大F.linear (CCCL histogram pattern已经实现了，但可以更激进)
|
||||
|
||||
### 修复3: computility-run.yaml 参数对齐
|
||||
|
||||
确保部署时真的用了yaml里的参数。max-num-seqs=2让n=2请求不会崩溃。
|
||||
|
||||
### 修复4: d01/d04速度
|
||||
|
||||
d01: 95.85s产了1085个token，约11.3 tok/s — 其实tps不太差
|
||||
对手d01: 8.49s产了139个token，约16.4 tok/s
|
||||
|
||||
**关键差异不是tps，是产了多少token!** 我们1085 vs 对手139。
|
||||
我们的模型在thinking里产了大量token。
|
||||
d01是basic_nostream，没有tool，所以enable_thinking=True是默认的。
|
||||
thinking产了1102个reasoning token + 0个content token。
|
||||
|
||||
**问题: d01测试的content[0]意味着没有实际内容输出!**
|
||||
对手content[11]说明他输出了内容。
|
||||
|
||||
这又回到了GatedDeltaNet NaN → 模型质量差的问题。
|
||||
|
||||
## 六、CCCL启示
|
||||
|
||||
CCCL在处理数值稳定性方面的核心设计:
|
||||
1. `overflow_cast_t<T>` — 在可能溢出的地方用更高精度的中间类型
|
||||
2. `cc_dispatch` — 不同硬件不同策略，不硬编码
|
||||
3. `policy_selector` — 基于benchmark数据选择参数，不拍脑袋
|
||||
|
||||
我们的DeltaNet实现缺少CCCL级别的数值稳定性保证。
|
||||
@@ -232,12 +232,18 @@ def _torch_chunk_gated_delta_rule(
|
||||
torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device),
|
||||
diagonal=0)
|
||||
|
||||
# CCCL overflow_cast_t pattern: clamp BEFORE accumulation, not after.
|
||||
# Without pre-clamp, cumsum of large g values produces huge numbers
|
||||
# that downstream exp() and matmul amplify into NaN.
|
||||
# BI-V100 docker logs show 99.98-100% NaN rate in every GatedDeltaNet layer.
|
||||
#
|
||||
# Pre-clamp: limit each g element so cumsum over chunk_size stays bounded.
|
||||
# With chunk_size=64 and per-element clamp ±0.3, cumsum range ≈ ±19.2.
|
||||
# Post-clamp to ±12 keeps exp(g_diff) ≤ exp(24) ≈ 2.6e10 — safe for
|
||||
# float32 matmul accumulation (k_dim=64 → max product ~1.7e12, within float32).
|
||||
g = g.clamp(-0.5, 0.5)
|
||||
g = g.cumsum(dim=-1)
|
||||
# Clamp gate logits to prevent exp overflow → NaN cascade.
|
||||
# CCCL dispatch_reduce_deterministic.cuh: numerical stability requires
|
||||
# bounded intermediate values. Gate logit range [-20, 20] keeps exp
|
||||
# in [~2e-9, ~5e8] — safe for float32 accumulation.
|
||||
g = g.clamp(-20.0, 20.0)
|
||||
g = g.clamp(-12.0, 12.0)
|
||||
decay_mask = ((g.unsqueeze(-1) - g.unsqueeze(-2)).tril().exp().float()).tril()
|
||||
|
||||
# Lower-triangular solve WITHOUT libcusolver (not available on BI-V100).
|
||||
@@ -269,16 +275,22 @@ def _torch_chunk_gated_delta_rule(
|
||||
return torch.linalg.solve_triangular(
|
||||
IminusA, rhs, upper=False, unitriangular=True)
|
||||
else:
|
||||
# Python forward substitution fallback
|
||||
# Python forward substitution fallback with numerical stability.
|
||||
# CCCL overflow_cast pattern: clamp intermediate results per row
|
||||
# to prevent the A @ x accumulation from amplifying small errors
|
||||
# into NaN. Without this, BI-V100 shows 100% NaN in every DeltaNet layer.
|
||||
x = torch.zeros_like(rhs)
|
||||
x[..., 0, :] = rhs[..., 0, :]
|
||||
x[..., 0, :] = rhs[..., 0, :].clamp(-1e4, 1e4)
|
||||
for i in range(1, C):
|
||||
x[..., i, :] = rhs[..., i, :] + (A_lower[..., i, :i].unsqueeze(-2) @ x[..., :i, :]).squeeze(-2)
|
||||
correction = (A_lower[..., i, :i].unsqueeze(-2) @ x[..., :i, :]).squeeze(-2)
|
||||
x[..., i, :] = (rhs[..., i, :] + correction).clamp(-1e4, 1e4)
|
||||
return x
|
||||
|
||||
value = _forward_sub_lower(A, v_beta)
|
||||
|
||||
k_cumdecay = _forward_sub_lower(A, k_beta * g.exp().unsqueeze(-1))
|
||||
# Clamp g.exp() to prevent k_cumdecay from having extreme values
|
||||
# that would amplify in the forward substitution loop.
|
||||
k_cumdecay = _forward_sub_lower(A, k_beta * g.exp().clamp(-1e4, 1e4).unsqueeze(-1))
|
||||
|
||||
del A # free memory
|
||||
|
||||
@@ -319,6 +331,10 @@ def _torch_chunk_gated_delta_rule(
|
||||
+ (k_i * g_diff_exp[:, :, i, :, None])
|
||||
.transpose(-1, -2) @ v_new
|
||||
)
|
||||
# CCCL numerical guard: clamp state to prevent cross-chunk accumulation
|
||||
# from amplifying into NaN. State elements represent k_dim × v_dim
|
||||
# attention memory; values beyond ±1e4 indicate numerical runaway.
|
||||
last_state = last_state.clamp(-1e4, 1e4)
|
||||
|
||||
if not output_final_state:
|
||||
last_state = None
|
||||
@@ -544,9 +560,13 @@ class GatedDeltaNet(nn.Module):
|
||||
|
||||
beta = b_all[s:e].sigmoid().unsqueeze(0) # (1, seq_len, local_num_v)
|
||||
# CCCL overflow_cast pattern: clamp before exp to prevent
|
||||
# overflow → NaN cascade. A_log.exp() can exceed float32 range
|
||||
# when A_log > ~88; clamping to [-20,20] keeps exp in safe range.
|
||||
_A_safe = self.A_log.float().clamp(-20.0, 20.0)
|
||||
# overflow → NaN cascade. Tightened to [-5,5] because:
|
||||
# A_log.exp() range [0.007, 148.4] — moderate decay rates.
|
||||
# Multiplied by softplus(a + dt_bias) ≈ [0.7, 10] → g ≈ [-1484, -0.005]
|
||||
# Per-element g then gets clamped to [-0.5, 0.5] in chunk_gated_delta_rule.
|
||||
# The tighter clamp here prevents A_log outliers from creating
|
||||
# extreme g values before the chunk-level clamp catches them.
|
||||
_A_safe = self.A_log.float().clamp(-5.0, 5.0)
|
||||
g = (-_A_safe.exp()
|
||||
* F.softplus(a_all[s:e].float() + self.dt_bias)
|
||||
).unsqueeze(0) # (1, seq_len, local_num_v)
|
||||
@@ -634,8 +654,8 @@ class GatedDeltaNet(nn.Module):
|
||||
v = v.reshape(num_seqs, 1, local_num_v, self.head_v_dim)
|
||||
|
||||
beta = b_all.sigmoid().unsqueeze(1) # (num_seqs, 1, local_num_v)
|
||||
# CCCL overflow_cast pattern: clamp before exp (same as prefill path)
|
||||
_A_safe = self.A_log.float().clamp(-20.0, 20.0)
|
||||
# CCCL overflow_cast pattern: tightened to [-5,5] matching prefill path
|
||||
_A_safe = self.A_log.float().clamp(-5.0, 5.0)
|
||||
g = (-_A_safe.exp()
|
||||
* F.softplus(a_all.float() + self.dt_bias)
|
||||
).unsqueeze(1) # (num_seqs, 1, local_num_v)
|
||||
@@ -654,7 +674,7 @@ class GatedDeltaNet(nn.Module):
|
||||
q_t = _l2norm(q.squeeze(1)).float() * _scale # (B, H_v, k_dim)
|
||||
k_t = _l2norm(k.squeeze(1)).float() # (B, H_v, k_dim)
|
||||
v_t = v.squeeze(1).float() # (B, H_v, v_dim)
|
||||
g_t = g.squeeze(1).float().clamp_(-20.0, 20.0).exp_() # (B, H_v) overflow_cast
|
||||
g_t = g.squeeze(1).float().clamp_(-12.0, 12.0).exp_() # (B, H_v) overflow_cast tightened
|
||||
bt = beta.squeeze(1).float() # (B, H_v)
|
||||
|
||||
# Decay state in-place: (B, H_v, k_dim, v_dim) *= scalar per head
|
||||
@@ -676,6 +696,8 @@ class GatedDeltaNet(nn.Module):
|
||||
k_t.view(BH, self.head_k_dim, 1),
|
||||
delta.view(BH, 1, self.head_v_dim),
|
||||
)
|
||||
# CCCL numerical guard: clamp decode state (same as prefill cross-chunk)
|
||||
ts_flat.clamp_(-1e4, 1e4)
|
||||
|
||||
# Output: core_out = q_t @ updated temporal_state
|
||||
core_out = torch.bmm(
|
||||
|
||||
Reference in New Issue
Block a user