From ff971686d4bb255911e525472d6b88beaa8df956 Mon Sep 17 00:00:00 2001
From: Claude <claude@anthropic.com>
Date: Sat, 8 Aug 2026 15:08:00 +0000
Subject: [PATCH] =?UTF-8?q?fix(CRITICAL):=20max=5Fmodel=5Flen=20100000?=
 =?UTF-8?q?=E2=86=9280000=20(KV=20cache=20only=2088112)=20+=20NaN=20fix?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Docker log proves two fatal issues:

1. max_model_len=100000 > KV cache capacity 88112 → ValueError crash
   'max seq len (100000) is larger than maximum number of tokens
    that can be stored in KV cache (88112)'
   Fix: set max_model_len=80000 (safe margin below 88112)

2. NaN in GatedDeltaNet layers 34,36,37,38 (frac=1.0000)
   Root cause: g.cumsum() → g.exp() overflow to inf → inf*0 = NaN
   Fix: clamp all g values to [-80,80] before exp() calls
   (max safe float32 exp input ~88, use 80 for margin)
   Applied to: cumsum result, k_cumdecay, attn_inter, last_state update

3. CoreX modules confirmed NOT in base image:
   'CoreX GDN module not found'
   'CoreX MoE module not found'
   → pure PyTorch is the only path, must be numerically stable
---
 computility-run.yaml       |  2 +-
 qwen3_6_scripts/qwen3_5.py | 11 +++++++----
 2 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/computility-run.yaml b/computility-run.yaml
index f88472f5..b03b3886 100644
--- a/computility-run.yaml
+++ b/computility-run.yaml
@@ -8,7 +8,7 @@ command:
     - --served-model-name
     - llm
     - --max-model-len
-    - '100000'
+    - '80000'
     - --gpu-memory-utilization
     - '0.9'
     - --trust-remote-code
diff --git a/qwen3_6_scripts/qwen3_5.py b/qwen3_6_scripts/qwen3_5.py
index 4b7124fa..45b8ffa9 100644
--- a/qwen3_6_scripts/qwen3_5.py
+++ b/qwen3_6_scripts/qwen3_5.py
@@ -144,6 +144,9 @@ def _torch_chunk_gated_delta_rule(
         diagonal=0)
 
     g = g.cumsum(dim=-1)
+    # Numerical stability: clamp cumulative decay to prevent exp() → inf → NaN
+    # CCCL pattern: block_scan overflow guard. Max safe float32 exp input ~88.
+    g = g.clamp(-80.0, 80.0)
     decay_mask = ((g.unsqueeze(-1) - g.unsqueeze(-2)).tril().exp().float()).tril()
     attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask_upper, 0)
     for i in range(1, chunk_size):
@@ -152,7 +155,7 @@ def _torch_chunk_gated_delta_rule(
         attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
     attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
     value = attn @ v_beta
-    k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1))
+    k_cumdecay = attn @ (k_beta * g.clamp(-80, 80).exp().unsqueeze(-1))
 
     last_state = (
         torch.zeros(batch, num_heads, k_dim, v_dim, dtype=value.dtype, device=value.device)
@@ -169,11 +172,11 @@ def _torch_chunk_gated_delta_rule(
         attn_i = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask_upper2, 0)
         v_prime = k_cumdecay[:, :, i] @ last_state
         v_new = v_i - v_prime
-        attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_state
+        attn_inter = (q_i * g[:, :, i, :, None].clamp(-80, 80).exp()) @ last_state
         core_out[:, :, i] = attn_inter + attn_i @ v_new
         last_state = (
-            last_state * g[:, :, i, -1, None, None].exp()
-            + (k_i * (g[:, :, i, -1, None] - g[:, :, i]).exp()[..., None])
+            last_state * g[:, :, i, -1, None, None].clamp(-80, 80).exp()
+            + (k_i * (g[:, :, i, -1, None] - g[:, :, i]).clamp(-80, 80).exp()[..., None])
             .transpose(-1, -2) @ v_new
         )