linoyts HF Staff commited on
Commit
4675f39
·
verified ·
1 Parent(s): 14f6b0b

Stream: render every edit (no debounce), 0.5 s segments, 0.6 s lead; reuse KV cache

Browse files
Files changed (2) hide show
  1. README.md +2 -2
  2. app.py +11 -7
README.md CHANGED
@@ -22,8 +22,8 @@ models:
22
  Real-time, prompt-reactive text-to-video with [apple/FastFrames-S](https://huggingface.co/apple/FastFrames-S).
23
  Press Go live and keep editing the prompt; the stream follows it.
24
 
25
- - One held ZeroGPU session per user streams 1 s segments; prompt edits reach it through a per-session file.
26
- - A prompt edit (debounced) regenerates the scene from frame 0 with the same seed, so the new prompt is followed exactly
27
  and the shot stays comparable.
28
  - An idle prompt keeps streaming past 5 s with a sink + rolling KV window (keys re-rotated so RoPE positions stay in
29
  0..20) and gets a fresh scene every 20 s (adjustable) to stop long-rollout drift.
 
22
  Real-time, prompt-reactive text-to-video with [apple/FastFrames-S](https://huggingface.co/apple/FastFrames-S).
23
  Press Go live and keep editing the prompt; the stream follows it.
24
 
25
+ - One held ZeroGPU session per user streams 0.5 s segments; prompt edits reach it through a per-session file.
26
+ - Every prompt edit regenerates the scene from frame 0 with the same seed, so the new prompt is followed exactly
27
  and the shot stays comparable.
28
  - An idle prompt keeps streaming past 5 s with a sink + rolling KV window (keys re-rotated so RoPE positions stay in
29
  0..20) and gets a fresh scene every 20 s (adjustable) to stop long-rollout drift.
app.py CHANGED
@@ -4,7 +4,7 @@ One held ZeroGPU session (like the Qwen-Image 2.1 Realtime Space) runs the offic
4
  per-frame rollout and streams it as 1 s segments. UI edits reach the running GPU process through a
5
  small per-session JSON file.
6
 
7
- * A prompt edit (debounced) regenerates the scene from frame 0 under the new prompt with the same
8
  seed, so the shot stays comparable and the new prompt is followed exactly.
9
  * While the prompt is idle the scene keeps going past the trained 5 s with a sink + rolling KV
10
  window (slot 0 keeps the first frame; when the 21-slot cache is full the oldest window frames are
@@ -41,12 +41,12 @@ WIDTH, HEIGHT = 832, 480
41
  MAX_SEED = 2**31 - 1
42
  MAX_CHARS = 1500
43
 
44
- CHUNK = int(os.environ.get("CHUNK", 4)) # latent frames per streamed segment (4 = 1 s of video)
45
  ROLL_STRIDE = int(os.environ.get("ROLL_STRIDE", 4)) # window frames dropped each time the cache is full
46
  SINK = os.environ.get("SINK", "1") == "1" # keep frame 0 as an attention sink
47
- LEAD = float(os.environ.get("LEAD", 1.25)) # max seconds of video generated ahead of playback
48
  REGEN_SECONDS = int(os.environ.get("REGEN_SECONDS", 20)) # idle prompt: fresh scene after this much video
49
- DEBOUNCE = float(os.environ.get("DEBOUNCE", 0.4)) # a prompt must stay unchanged this long before it applies
50
  GPU_SIZE = os.environ.get("GPU_SIZE", "xlarge")
51
  SESSION_SECONDS = int(os.environ.get("SESSION_SECONDS", 90))
52
  RT_WINDOW = SESSION_SECONDS - 6 # leave headroom under the @spaces.GPU duration
@@ -75,8 +75,10 @@ print(f"[load] FastFrames-S ready: tokens/frame={TOKENS} cap={CAP} temporal_pair
75
  class LiveRollout:
76
  """FastFrames.rollout, one latent frame at a time, forever, with a sink + rolling KV window."""
77
 
78
- def __init__(self, seed):
79
- self.kv = allocate_kv_cache(PIPE.dit, CAP, TOKENS, device=PIPE.device)
 
 
80
  self.gen = torch.Generator(device=PIPE.device).manual_seed(int(seed))
81
  self.used = 0 # occupied cache slots
82
  self.total = 0 # latent frames generated
@@ -254,9 +256,11 @@ def live_session(sid):
254
  scenes = 0
255
  last_ver = -1
256
 
 
 
257
  def new_scene(reason):
258
  nonlocal rollout, scenes
259
- rollout = LiveRollout(seed)
260
  scenes += 1
261
  print(f"[live] {sid[:6]} scene {scenes} ({reason}) seed={seed} {prompt[:60]!r}", flush=True)
262
 
 
4
  per-frame rollout and streams it as 1 s segments. UI edits reach the running GPU process through a
5
  small per-session JSON file.
6
 
7
+ * Every prompt edit regenerates the scene from frame 0 under the new prompt with the same
8
  seed, so the shot stays comparable and the new prompt is followed exactly.
9
  * While the prompt is idle the scene keeps going past the trained 5 s with a sink + rolling KV
10
  window (slot 0 keeps the first frame; when the 21-slot cache is full the oldest window frames are
 
41
  MAX_SEED = 2**31 - 1
42
  MAX_CHARS = 1500
43
 
44
+ CHUNK = int(os.environ.get("CHUNK", 2)) # latent frames per streamed segment (2 = 0.5 s of video)
45
  ROLL_STRIDE = int(os.environ.get("ROLL_STRIDE", 4)) # window frames dropped each time the cache is full
46
  SINK = os.environ.get("SINK", "1") == "1" # keep frame 0 as an attention sink
47
+ LEAD = float(os.environ.get("LEAD", 0.6)) # max seconds of video generated ahead of playback
48
  REGEN_SECONDS = int(os.environ.get("REGEN_SECONDS", 20)) # idle prompt: fresh scene after this much video
49
+ DEBOUNCE = float(os.environ.get("DEBOUNCE", 0.0)) # a prompt must stay unchanged this long before it applies
50
  GPU_SIZE = os.environ.get("GPU_SIZE", "xlarge")
51
  SESSION_SECONDS = int(os.environ.get("SESSION_SECONDS", 90))
52
  RT_WINDOW = SESSION_SECONDS - 6 # leave headroom under the @spaces.GPU duration
 
75
  class LiveRollout:
76
  """FastFrames.rollout, one latent frame at a time, forever, with a sink + rolling KV window."""
77
 
78
+ def __init__(self, seed, kv):
79
+ # one cache per session, reused across scenes: slots are overwritten in order and attention only
80
+ # reads filled slots, so no clearing is needed
81
+ self.kv = kv
82
  self.gen = torch.Generator(device=PIPE.device).manual_seed(int(seed))
83
  self.used = 0 # occupied cache slots
84
  self.total = 0 # latent frames generated
 
256
  scenes = 0
257
  last_ver = -1
258
 
259
+ kv = allocate_kv_cache(PIPE.dit, CAP, TOKENS, device=PIPE.device)
260
+
261
  def new_scene(reason):
262
  nonlocal rollout, scenes
263
+ rollout = LiveRollout(seed, kv)
264
  scenes += 1
265
  print(f"[live] {sid[:6]} scene {scenes} ({reason}) seed={seed} {prompt[:60]!r}", flush=True)
266