Replace sfp4 checkpoint-750 with checkpoint-700

README.md

@@ -1,9 +1,9 @@
-# sfp4_v4_sparse09_hpo_on_ours_p_init2050 checkpoint-750
+# sfp4_v4_sparse09_hpo_on_ours_p_init2050 checkpoint-700
 
 This upload contains the consolidated WanTransformer3DModel transformer weights
 from:
 
-`checkpoints/sfp4_v4_sparse09_hpo_on_ours_p_init2050_1n_interactive/checkpoint-750`
+`checkpoints/sfp4_v4_sparse09_hpo_on_ours_p_init2050_1n_interactive/checkpoint-700`
 
 Contents:
 
@@ -30,7 +30,7 @@ SFT launch scripts.
 
 It also includes the inference entrypoint snapshot and an example script:
 
-- `backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_750.sh`
+- `backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_700.sh`
 - `backend_snapshot/training_attention_settings.json`
 
 Attention setup for this checkpoint:

backend_snapshot/README.md

@@ -1,8 +1,8 @@
-# Backend snapshot for checkpoint-750
+# Backend snapshot for checkpoint-700
 
 This directory is the code snapshot for the training backend used by:
 
-`sfp4_v4_sparse09_hpo_on_ours_p_init2050_1n_interactive/checkpoint-750`
+`sfp4_v4_sparse09_hpo_on_ours_p_init2050_1n_interactive/checkpoint-700`
 
 Key runtime settings:
 
@@ -38,7 +38,7 @@ Important files:
 - `scripts/training/run_sparse_fp4_train_v4_common.sh`: common SFT launch/resume script.
 - `training_attention_settings.json`: structured attention/training settings
   for this checkpoint.
-- `scripts/inference/run_sfp4_ours_p_checkpoint_750.sh`: inference example
+- `scripts/inference/run_sfp4_ours_p_checkpoint_700.sh`: inference example
   for the uploaded transformer checkpoint.
 - `fastvideo/entrypoints/cli/generate.py`, `fastvideo/entrypoints/video_generator.py`,
   `fastvideo/pipelines/basic/wan/wan_pipeline.py`, and
@@ -53,7 +53,7 @@ hf download yitongl/sparse_quant_exp \
   --local-dir checkpoints/hf_download/sparse_quant_exp \
   --include 'transformer/*'
 
-bash backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_750.sh
+bash backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_700.sh
 ```
 
 Source repo HEAD when staged:

backend_snapshot/manifest.sha256

@@ -1,4 +1,4 @@
-033b7cce6eb0ead450a35b910adc9ae1323b8d2323aea6731b72b940e222fb46  ./README.md
+936c4ab834574320b4cca77fe95ffea436c15c6e49b0ad4bd21d8ebf32701fc8  ./README.md
 9c7dec8f1b8160954d0566231b0952a5f6a5d81f546affd71d190b2b3fc79cb6  ./examples/inference/basic/basic.py
 9d1d8dc58aab529270fe31eb1735d6a1382c0c6d36fccca122a8dbffa1b714fd  ./fastvideo-kernel/python/fastvideo_kernel/block_sparse_attn_ours_p.py
 211c7f0445fbe9488250f01fa83457c6620e83bd6f3877db791fd155de93c08b  ./fastvideo-kernel/python/fastvideo_kernel/triton_kernels/block_sparse_attn_triton_ours_p.py
@@ -25,9 +25,9 @@ c046b1914041b59254bcdfe577aed20d6f007a72632ea1fe1ae92fa678eca760  ./fastvideo/pl
 2456d39ca28019e12bb7ab007774e86348f0582a017bf0e6c91e2a01d654a1a0  ./fastvideo/train/models/wan/wan.py
 bc46e84b732567de6c0325223405daecd1226c623e303be33c7be9b5b7fdec08  ./fastvideo/training/training_pipeline.py
 1d3898fa37e21029df6c37e05dc34ed7805a211c2f87de6642db890e5a8c6f2e  ./fastvideo/training/wan_training_pipeline.py
-57fd33c78a16c9b4a239734c76726d94df1daf86f15cf22451c6107fcc197834  ./scripts/inference/run_sfp4_ours_p_checkpoint_750.sh
+9689cd17d841b713bb1627277e63bf436ac9a0a2f62f3f1ed7f9a6462c4def1a  ./scripts/inference/run_sfp4_ours_p_checkpoint_700.sh
 0162b26dddb2a249e4e2cc56a7a28eba4fb3ea77e938eaf9c6b16a241edcd3ec  ./scripts/inference/run_sfp4_single.sh
 159579109f9fb7d7dded977f5c7fc974583c95f3f895e418d6c466463c036304  ./scripts/inference/run_validate_and_gen.sh
 5c982b64653fae83ebfdeb43fda8f29b3e2cb581fb4daee38cd3cf56aa9d73f5  ./scripts/training/run_sparse_fp4_train_v4_1n_sparse09_hpo_on_ours_p_init2050_interactive.sh
 5c1d5ce9ecc8b90e59ddfc2ddb3e2dae500bcd3acb90429c901444b1630f05fb  ./scripts/training/run_sparse_fp4_train_v4_common.sh
-75455829ca55a80daaa7e3c7faa080b6eec3c7109bdbd1e198b722face62eed0  ./training_attention_settings.json
+8ddeea65247d9fa31a4a8a2a5ce5abe068a911ff4d67871453555e1355af8ecf  ./training_attention_settings.json

backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_700.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+
+set -euo pipefail
+
+REPO_ROOT="${REPO_ROOT:-/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/yitongl/code/FastVideo}"
+MODEL_PATH="${MODEL_PATH:-Wan-AI/Wan2.1-T2V-1.3B-Diffusers}"
+HF_REPO="${HF_REPO:-yitongl/sparse_quant_exp}"
+CHECKPOINT_DIR="${CHECKPOINT_DIR:-${REPO_ROOT}/checkpoints/hf_download/sparse_quant_exp}"
+WEIGHTS_PATH="${WEIGHTS_PATH:-${CHECKPOINT_DIR}/transformer/diffusion_pytorch_model.safetensors}"
+OUTPUT_DIR="${OUTPUT_DIR:-${REPO_ROOT}/outputs/sfp4_v4_sparse09_checkpoint_700}"
+PROMPT="${PROMPT:-A cinematic shot of a futuristic city street at dusk, reflective pavement, soft volumetric light, detailed motion, stable camera.}"
+NEGATIVE_PROMPT="${NEGATIVE_PROMPT:-Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards}"
+
+cd "${REPO_ROOT}"
+
+if [[ -f ".venv/bin/activate" ]]; then
+    source .venv/bin/activate
+fi
+
+export PYTHONPATH="${REPO_ROOT}/fastvideo-kernel/python:${REPO_ROOT}/fastvideo-kernel:${PYTHONPATH:-}"
+export FASTVIDEO_ATTENTION_BACKEND=SPARSE_FP4_OURS_P_ATTN
+export FASTVIDEO_SPARSE_FP4_USE_HIGH_PREC_O=1
+
+if [[ ! -f "${WEIGHTS_PATH}" ]]; then
+    echo "Missing ${WEIGHTS_PATH}"
+    echo "Download the uploaded transformer weights first:"
+    echo "  hf download ${HF_REPO} --repo-type model --local-dir ${CHECKPOINT_DIR} --include 'transformer/*'"
+    exit 1
+fi
+
+mkdir -p "${OUTPUT_DIR}"
+
+fastvideo generate \
+    --model-path "${MODEL_PATH}" \
+    --init-weights-from-safetensors "${WEIGHTS_PATH}" \
+    --sp-size 1 \
+    --tp-size 1 \
+    --num-gpus 1 \
+    --dit-cpu-offload False \
+    --vae-cpu-offload False \
+    --text-encoder-cpu-offload True \
+    --pin-cpu-memory False \
+    --height 448 \
+    --width 832 \
+    --num-frames 77 \
+    --num-inference-steps 50 \
+    --fps 16 \
+    --guidance-scale 5.0 \
+    --flow-shift 1.0 \
+    --prompt "${PROMPT}" \
+    --negative-prompt "${NEGATIVE_PROMPT}" \
+    --seed 1000 \
+    --VSA-sparsity 0.9 \
+    --output-path "${OUTPUT_DIR}/"

backend_snapshot/training_attention_settings.json

@@ -1,6 +1,6 @@
 {
   "run_name": "sfp4_v4_sparse09_hpo_on_ours_p_init2050_1n_interactive",
-  "checkpoint": "checkpoint-750",
+  "checkpoint": "checkpoint-700",
   "training_method": "legacy_sft_wan_training_pipeline",
   "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
   "init_weights_from_safetensors": "checkpoints/init/sfp4_v4_sparse06_hpo_on_ours_p_1n_interactive_v2_ckpt2050/transformer/diffusion_pytorch_model.safetensors",

transformer/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:18b2922bb6e0480753e63da2488b6dd1f68cd23e6e2257e4007295d3a2ea5e0a
+oid sha256:4595ca81ea7085c15ccf14b738aa9c0fdf2d2786641f49b55e0bc0e99bf042d2
 size 5676070784