wx44wx/icassp2027-experiment
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add causal MiniCPM-o streaming feature extraction

Browse Source
This commit is contained in:
Xin Wang
2026-09-17 01:08:50 +08:00
parent 45facc8bf4
commit 6d0c8b733b
5 changed files with 1533 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

204
docs/minicpmo_streaming_features.md Normal file
View File

@@ -0,0 +1,204 @@
# MiniCPM-o 4.5 causal global features
`streaming_minicpmo_features_directly.py` uses MiniCPM-o 4.5 only as a frozen
audio--visual feature extractor. It never calls `streaming_generate`, never
loads TTS, and never lets model-generated speech/listen tokens enter the
conversation history.
## Frozen modes
### `native-1hz-hold`
- The model commits one input-only AV chunk per completed second.
- The newest completed state is forward-held onto the 5 Hz target grid.
- The first four rows (0.2--0.8 s) use an all-zero initialization. The 1.0-s
state is never backfilled into those rows.
- `state_source_times`, `native_update_mask`, and `initialization_mask` make
this alignment auditable.
### `cached-partial-5hz`
- Completed one-second chunks remain in the official MiniCPM streaming cache.
- At a non-boundary 200-ms cutoff, the extractor snapshots the LLM cache,
audio-encoder cache, streaming Mel buffer, chunk index, and RNG state through
MiniCPM's official speculative-state API.
- It appends only the current half-open AV tail, reads the final normalized
Qwen3 hidden state, and restores the snapshot. Temporary closure never enters
the main cache.
- Native commits retain MiniCPM's normal streaming-window policy. Destructive
round pruning is suspended only inside the temporary transaction because the
upstream snapshot can truncate appended KV tokens but cannot recreate a
previously pruned round.
- At a one-second boundary, a complete native chunk is committed.
The 5 Hz branch is saved as the normal `mode=causal-prefix` contract only after
all mandatory checks pass:
1. a temporary full chunk matches the native committed state at a one-second
boundary;
2. a clean stream replay matches the cached result;
3. changing audio strictly after the cutoff leaves the current state exactly
unchanged;
4. snapshot restoration leaves LLM/audio cache lengths and checksums, chunk
history, position offset, and the audio chunk index unchanged.
With `--allow-unverified-partial-tail`, a failed or too-short audit may be
saved only as `mode=minicpmo-unverified-partial-tail`; the default head loader
rejects that artifact.
## Environment
The inspected upstream revision requires Python 3.10 and
`transformers==4.51.0`. Use a dedicated environment rather than changing the
working Qwen environment.
```bash
python -m pip install \
"transformers==4.51.0" \
accelerate \
"torch>=2.3.0,<=2.8.0" \
"torchaudio<=2.8.0" \
"minicpmo-utils>=1.0.5" \
"qwen-omni-utils[decord]"
```
The commands below use the upstream revision inspected during implementation,
`1f761131fa83f5ed3cd6f2f22b225c4501d154fa`. If the model is already stored
locally, replace `--model-id` with that snapshot directory and record its
immutable revision separately.
## Required smoke sequence
First validate the source timeline and decoder without loading MiniCPM:
```bash
python streaming_minicpmo_features_directly.py \
--easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
--end 3 \
--mode native-1hz-hold \
--decoder decord \
--media-only
```
Then extract three seconds with the native held-state baseline:
```bash
CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
--model-id openbmb/MiniCPM-o-4_5 \
--revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
--easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
--end 3 \
--mode native-1hz-hold \
--dtype bfloat16 \
--attention sdpa \
--device cuda:0 \
--decoder decord \
--decoder-threads 1 \
--max-open-readers 1 \
--output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/smoke/Session_4_native_3s.pt
```
Run the verified cached partial-tail smoke test separately:
```bash
CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
--model-id openbmb/MiniCPM-o-4_5 \
--revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
--easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
--end 3 \
--mode cached-partial-5hz \
--verify-every 1 \
--future-intervention-every 1 \
--dtype bfloat16 \
--attention sdpa \
--device cuda:0 \
--decoder decord \
--decoder-threads 1 \
--max-open-readers 1 \
--output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/smoke/Session_4_partial_3s.pt
```
Do not add `--allow-unverified-partial-tail` to a paper run. A verification
failure is a protocol failure to diagnose, not a tolerance to bypass.
Inspect both artifacts before a full session:
```bash
python - <<'PY'
from pathlib import Path
import torch
root = Path("/data1/yfzx/wangx/Datasets/easycom/minicpmo_features/smoke")
for name in ("Session_4_native_3s.pt", "Session_4_partial_3s.pt"):
payload = torch.load(root / name, map_location="cpu", weights_only=False)
print(name)
print(" mode:", payload["mode"], payload["extractor_mode"])
print(" states:", tuple(payload["states"].shape))
print(" times:", payload["times"][:6].tolist())
print(" source times:", payload["state_source_times"][:6].tolist())
print(" verified:", payload["partial_tail_admissible"])
print(" checks:", payload["verification"])
PY
```
Expected alignment:
```text
native times: 0.2 0.4 0.6 0.8 1.0 1.2 ...
native source time: 0.0 0.0 0.0 0.0 1.0 1.0 ...
partial source time: 0.2 0.4 0.6 0.8 1.0 1.2 ...
```
## Full Session 4
Native 1 Hz held features:
```bash
mkdir -p /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/native_1hz_hold
CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
--model-id openbmb/MiniCPM-o-4_5 \
--revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
--easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
--target /data1/yfzx/wangx/Datasets/easycom/easycom_target_frame/Session_4.npy \
--mode native-1hz-hold \
--dtype bfloat16 \
--attention sdpa \
--device cuda:0 \
--decoder decord \
--decoder-threads 1 \
--max-open-readers 1 \
--frame-cache-size 32 \
--output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/native_1hz_hold/Session_4.pt
```
Verified 5 Hz cached partial-tail features:
```bash
mkdir -p /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/cached_partial_5hz
CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
--model-id openbmb/MiniCPM-o-4_5 \
--revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
--easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
--target /data1/yfzx/wangx/Datasets/easycom/easycom_target_frame/Session_4.npy \
--mode cached-partial-5hz \
--dtype bfloat16 \
--attention sdpa \
--device cuda:0 \
--decoder decord \
--decoder-threads 1 \
--max-open-readers 1 \
--frame-cache-size 32 \
--output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/cached_partial_5hz/Session_4.pt
```
The full run still executes the first clean-replay and future-intervention
checks plus every one-second boundary check. Add `--verify-every 250` and
`--future-intervention-every 250` for periodic 50-second audits; smaller values
make long sessions increasingly expensive because a clean audit replays the
prefix from the session start.
After extraction, either directory can be passed directly as
`train_turn_head.py --feature-root ...` because verified artifacts retain the
existing dense global-feature contract.