From d9f8e0d758351963585328d3cf9b26f8b33425eb Mon Sep 17 00:00:00 2001
From: "Tessa (livepeer-tessa)" <tessa@livepeer.org>
Date: Sun, 12 Apr 2026 18:21:15 +0000
Subject: [PATCH] fix(longlive): add logger + warning when KV cache write is
 skipped (#921)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The `if write_len > 0` guards in `CausalWanSelfAttention.forward()` already
prevent the RuntimeError (empty tensor slice when write_start_index ==
local_end_index), but the skip was silent, making it impossible to diagnose
cache-state drift from Grafana logs.

Changes:
- Add `import logging` and `logger = logging.getLogger(__name__)`
- Add `else: logger.warning(...)` branch on both the roll_and_insert and
  direct_insert write paths so that stale-index events are surfaced as
  WARNING-level log entries (rather than ~310 silent skips/session)

The WARNING includes write_start_index, local_end_index, and a reference to
#921 to aid future investigation of cache index drift after rapid
video↔text mode transitions.

Fixes daydreamlive/scope#921

Signed-off-by: Tessa (livepeer-tessa) <tessa@livepeer.org>
---
 .../longlive/modules/causal_model.py          | 22 +++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/src/scope/core/pipelines/longlive/modules/causal_model.py b/src/scope/core/pipelines/longlive/modules/causal_model.py
index 8607855bc..96e731484 100644
--- a/src/scope/core/pipelines/longlive/modules/causal_model.py
+++ b/src/scope/core/pipelines/longlive/modules/causal_model.py
@@ -1,5 +1,6 @@
 # Modified from https://github.com/NVlabs/LongLive
 # SPDX-License-Identifier: CC-BY-NC-SA-4.0
+import logging
 import math
 
 import torch
@@ -13,6 +14,7 @@
 )
 
 from scope.core.pipelines.wan2_1.modules.attention import attention
+
 from .model import (
     WAN_CROSSATTENTION_CLASSES,
     MLPProj,
@@ -30,6 +32,8 @@
     flex_attention, dynamic=False, mode="max-autotune-no-cudagraphs"
 )
 
+logger = logging.getLogger(__name__)
+
 
 def causal_rope_apply(x, grid_sizes, freqs, start_frame=0):
     n, c = x.size(2), x.size(3) // 2
@@ -331,6 +335,15 @@ def qkv_fn(x):
                     temp_v[:, write_start_index:local_end_index] = v[
                         :, roped_offset : roped_offset + write_len
                     ]
+                else:
+                    logger.warning(
+                        "KV cache write skipped (roll_and_insert): "
+                        "write_start_index (%d) >= local_end_index (%d) — "
+                        "cache indices may be stale after a rapid mode transition. "
+                        "(See daydreamlive/scope#921)",
+                        write_start_index,
+                        local_end_index,
+                    )
 
                 # Save cache update info for later use
                 cache_update_info = {
@@ -377,6 +390,15 @@ def qkv_fn(x):
                     temp_v[:, write_start_index:local_end_index] = v[
                         :, roped_offset : roped_offset + write_len
                     ]
+                else:
+                    logger.warning(
+                        "KV cache write skipped (direct_insert): "
+                        "write_start_index (%d) >= local_end_index (%d) — "
+                        "cache indices may be stale after a rapid mode transition. "
+                        "(See daydreamlive/scope#921)",
+                        write_start_index,
+                        local_end_index,
+                    )
 
                 # Save cache update info for later use
                 cache_update_info = {