File size: 12,132 Bytes
bdce33b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
diff --git a/generate_infinitetalk.py b/generate_infinitetalk.py
index e83daa8..77426ce 100644
--- a/generate_infinitetalk.py
+++ b/generate_infinitetalk.py
@@ -20,7 +20,10 @@ import wan
 from wan.configs import SIZE_CONFIGS, SUPPORTED_SIZES, WAN_CONFIGS
 from wan.utils.utils import str2bool, is_video, split_wav_librosa
 from wan.utils.multitalk_utils import save_video_ffmpeg
-from kokoro import KPipeline
+try:
+    from kokoro import KPipeline
+except ImportError:  # TTS stack unused when driving from a local audio file
+    KPipeline = None
 from transformers import Wav2Vec2FeatureExtractor
 from src.audio_analysis.wav2vec2 import Wav2Vec2Model
 from wan.utils.segvideo import shot_detect
@@ -544,6 +547,89 @@ def generate(args):
             num_persistent_param_in_dit=args.num_persistent_param_in_dit
         )
     
+    if os.environ.get("WAN_TRT") == "1":
+        # The engines own the transformer stack now: keep the PyTorch blocks off
+        # the GPU (~27GB) so the 18.6GB of engines fit alongside the rest.
+        import torch as _t
+        _n = len(wan_i2v.model.blocks)
+        wan_i2v.model.blocks = _t.nn.ModuleList([])   # the engines own these now
+        _t.cuda.empty_cache()
+        logging.info(f"WAN_TRT: dropped {_n} PyTorch blocks (~27GB); engines own the stack")
+
+    if os.environ.get("CAPTURE") == "1":
+        import sys as _sys
+        _sys.path.insert(0, "/workspace/trt")
+        from capture_calib import install_capture
+        install_capture(wan_i2v.model)
+
+    # --- FP8 (H100 native) + torch.compile toggles, applied after LoRA merge ---
+    if os.environ.get("WAN_FP8") == "1":
+        import torch as _t
+        from torchao.quantization import quantize_, Float8DynamicActivationFloat8WeightConfig
+        _dit = wan_i2v.model
+        # Only the transformer blocks: the tiny embedders/head stay bf16.
+        _n = 0
+        for _blk in _dit.blocks:
+            quantize_(_blk, Float8DynamicActivationFloat8WeightConfig())
+            _n += 1
+        logging.info(f"FP8: quantized {_n} transformer blocks (fp8 dynamic act + fp8 weight)")
+
+    if os.environ.get("WAN_COMPILE") == "1":
+        import torch as _t
+        _dit = wan_i2v.model
+        for _i, _blk in enumerate(_dit.blocks):
+            _dit.blocks[_i] = _t.compile(_blk, dynamic=False)
+        logging.info(f"compiled {len(_dit.blocks)} transformer blocks")
+
+    # --- profiling hook: set PROFILE=1 to time every DiT forward ---
+    if os.environ.get("PROFILE") == "1":
+        import time as _time, atexit as _atexit
+        import torch as _torch
+        _stats = {"n": 0, "t": 0.0}
+        _wall0 = _time.perf_counter()
+        _orig_fwd = wan_i2v.model.forward
+
+        _deep = os.environ.get("PROFILE_DEEP") == "1"
+
+        def _timed_fwd(*a, **kw):
+            _torch.cuda.synchronize()
+            _t0 = _time.perf_counter()
+            # kernel-level trace of a single steady-state forward (the 2nd)
+            if _deep and _stats["n"] == 1:
+                from torch.profiler import profile as _tp, ProfilerActivity as _PA
+                with _tp(activities=[_PA.CPU, _PA.CUDA], record_shapes=False) as _prof:
+                    out = _orig_fwd(*a, **kw)
+                    _torch.cuda.synchronize()
+                print("\n======= TOP CUDA KERNELS (one forward) =======")
+                print(_prof.key_averages().table(
+                    sort_by="self_cuda_time_total", row_limit=28,
+                    max_name_column_width=55))
+            else:
+                out = _orig_fwd(*a, **kw)
+            _torch.cuda.synchronize()
+            _stats["t"] += _time.perf_counter() - _t0
+            _stats["n"] += 1
+            return out
+
+        wan_i2v.model.forward = _timed_fwd
+
+        def _report():
+            n, t = _stats["n"], _stats["t"]
+            wall = _time.perf_counter() - _wall0
+            peak = _torch.cuda.max_memory_allocated() / 1e9
+            print("\n================ PROFILE ================")
+            print(f"DiT forwards        : {n}")
+            print(f"DiT total time      : {t:.1f} s")
+            if n:
+                print(f"DiT per forward     : {t / n * 1000:.0f} ms")
+            print(f"wall (post-load)    : {wall:.1f} s")
+            if wall > 0:
+                print(f"DiT share of wall   : {t / wall * 100:.0f} %")
+            print(f"peak VRAM allocated : {peak:.1f} GB")
+            print("=========================================")
+
+        _atexit.register(_report)
+
     generated_list = []
     with open(args.input_json, 'r', encoding='utf-8') as f:
         input_data = json.load(f)
diff --git a/wan/modules/multitalk_model.py b/wan/modules/multitalk_model.py
index 958e930..6eef6f3 100644
--- a/wan/modules/multitalk_model.py
+++ b/wan/modules/multitalk_model.py
@@ -21,6 +21,29 @@ try:
 except:
     USE_SAGEATTN = False
 
+from torch.nn.attention import sdpa_kernel, SDPBackend
+
+# cuDNN's Hopper attention beat sageattn 44.6ms vs 58.5ms at our shape.
+USE_CUDNN_SDPA = os.environ.get("WAN_CUDNN_ATTN", "1") == "1"
+
+# --- FP8 TensorRT block stack (WAN_TRT=1) -------------------------------------
+_TRT_STACK = None
+_TRT_SEQ_LEN = int(os.environ.get("WAN_TRT_SEQ_LEN", "30576"))   # 81 frames @ 448x832
+
+
+def _trt_stack():
+    """Lazily load the engines; returns None when TRT is off."""
+    global _TRT_STACK
+    if os.environ.get("WAN_TRT") != "1":
+        return None
+    if _TRT_STACK is None:
+        import sys
+        sys.path.insert(0, "/workspace/trt")
+        from trt_runner import TRTStack
+        _TRT_STACK = TRTStack()
+    return _TRT_STACK
+
+
 __all__ = ['WanModel']
 
 
@@ -50,30 +73,52 @@ def rope_params(max_seq_len, dim, theta=10000):
     return freqs
 
 
+_ROPE_CACHE = {}
+
+
+def _rope_cos_sin(grid_sizes, freqs, device, dtype):
+    """cos/sin RoPE table for this latent grid, built once and kept on-device."""
+    f, h, w = (int(v) for v in grid_sizes[0].tolist())
+    key = (f, h, w, str(device), dtype)
+    hit = _ROPE_CACHE.get(key)
+    if hit is not None:
+        return hit
+
+    c = freqs.size(1)
+    fr = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)
+    freqs_i = torch.cat([
+        fr[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
+        fr[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
+        fr[2][:w].view(1, 1, w, -1).expand(f, h, w, -1),
+    ], dim=-1).reshape(f * h * w, 1, -1)          # [L, 1, C/2] complex
+
+    cos = freqs_i.real.to(device=device, dtype=dtype).contiguous()
+    sin = freqs_i.imag.to(device=device, dtype=dtype).contiguous()
+    _ROPE_CACHE[key] = (cos, sin)
+    return cos, sin
+
+
 @amp.autocast(enabled=False)
 def rope_apply(x, grid_sizes, freqs):
-    s, n, c = x.size(1), x.size(2), x.size(3) // 2
+    """Real-valued RoPE. Mathematically identical to the complex fp64 version,
+    minus the per-layer host->device copy and the float64 traffic."""
+    b, s, n, d = x.shape
+    cos, sin = _rope_cos_sin(grid_sizes, freqs, x.device, torch.float32)
+    L = cos.size(0)
 
-    freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)
+    xf = x[:, :L].float().reshape(b, L, n, d // 2, 2)
+    x_r, x_i = xf[..., 0], xf[..., 1]
 
-    output = []
-    for i, (f, h, w) in enumerate(grid_sizes.tolist()):
-        seq_len = f * h * w
+    cos_ = cos.unsqueeze(0)                       # [1, L, 1, C/2]
+    sin_ = sin.unsqueeze(0)
 
-        x_i = torch.view_as_complex(x[i, :s].to(torch.float64).reshape(
-            s, n, -1, 2))
-        freqs_i = torch.cat([
-            freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
-            freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
-            freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
-        ],
-                            dim=-1).reshape(seq_len, 1, -1)
-        freqs_i = freqs_i.to(device=x_i.device)
-        x_i = torch.view_as_real(x_i * freqs_i).flatten(2)
-        x_i = torch.cat([x_i, x[i, seq_len:]])
+    o_r = x_r * cos_ - x_i * sin_
+    o_i = x_r * sin_ + x_i * cos_
+    out = torch.stack([o_r, o_i], dim=-1).flatten(3)
 
-        output.append(x_i)
-    return torch.stack(output).float()
+    if L < s:                                     # keep any padding untouched
+        out = torch.cat([out, x[:, L:].float()], dim=1)
+    return out.float()
 
 
 class WanRMSNorm(nn.Module):
@@ -137,7 +182,7 @@ class WanSelfAttention(nn.Module):
         self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
         self.norm_k = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
 
-    def forward(self, x, seq_lens, grid_sizes, freqs, ref_target_masks=None):
+    def forward(self, x, seq_lens, grid_sizes, freqs, ref_target_masks=None, human_num=None):
         b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
 
         # query, key, value function
@@ -151,7 +196,15 @@ class WanSelfAttention(nn.Module):
         q = rope_apply(q, grid_sizes, freqs)
         k = rope_apply(k, grid_sizes, freqs)
 
-        if USE_SAGEATTN:
+        if USE_CUDNN_SDPA:
+            # [B, L, H, D] -> [B, H, L, D] for SDPA, and back
+            qb = q.transpose(1, 2).to(torch.bfloat16)
+            kb = k.transpose(1, 2).to(torch.bfloat16)
+            vb = v.transpose(1, 2).to(torch.bfloat16)
+            with sdpa_kernel(SDPBackend.CUDNN_ATTENTION):
+                x = F.scaled_dot_product_attention(qb, kb, vb)
+            x = x.transpose(1, 2).type_as(v)
+        elif USE_SAGEATTN:
             x = sageattn(q.to(torch.bfloat16), k.to(torch.bfloat16), v, tensor_layout='NHD')
         else:
             x = flash_attention(
@@ -165,9 +218,15 @@ class WanSelfAttention(nn.Module):
         # output
         x = x.flatten(2)
         x = self.o(x)
-        with torch.no_grad():
-            x_ref_attn_map = get_attn_map_with_target(q.type_as(x), k.type_as(x), grid_sizes[0], 
-                                                    ref_target_masks=ref_target_masks)
+        # The ref-attn map only feeds SingleStreamMutiAttention's multi-speaker routing;
+        # with one speaker that branch short-circuits and never reads it, so skip building
+        # a [heads, seq, ref_seq] map (GBs) in every layer.
+        if human_num == 1:
+            x_ref_attn_map = None
+        else:
+            with torch.no_grad():
+                x_ref_attn_map = get_attn_map_with_target(q.type_as(x), k.type_as(x), grid_sizes[0],
+                                                        ref_target_masks=ref_target_masks)
 
         return x, x_ref_attn_map
 
@@ -294,7 +353,7 @@ class WanAttentionBlock(nn.Module):
         # self-attention
         y, x_ref_attn_map = self.self_attn(
             (self.norm1(x).float() * (1 + e[1]) + e[0]).type_as(x), seq_lens, grid_sizes,
-            freqs, ref_target_masks=ref_target_masks)
+            freqs, ref_target_masks=ref_target_masks, human_num=human_num)
         with amp.autocast(dtype=torch.float32):
             x = x + y * e[2]
         
@@ -757,7 +816,14 @@ class WanModel(ModelMixin, ConfigMixin):
                     for block in self.blocks:
                         x = block(x, **kwargs)
                     self.previous_residual_uncond = x - ori_x
+        elif _trt_stack() is not None and x.shape[1] == _TRT_SEQ_LEN:
+            cos, sin = _rope_cos_sin(grid_sizes, self.freqs, x.device, torch.float32)
+            x = _trt_stack()(x, e0, context, audio_embedding.squeeze(0), cos, sin)
         else:
+            if os.environ.get("WAN_TRT") == "1":
+                raise RuntimeError(
+                    f"WAN_TRT=1 but seq_len {x.shape[1]} != engine seq_len {_TRT_SEQ_LEN}. "
+                    f"Engines are static; rebuild them for this frame count/resolution.")
             for block in self.blocks:
                 x = block(x, **kwargs)
 
diff --git a/wan/multitalk.py b/wan/multitalk.py
index be7819e..1e9513b 100644
--- a/wan/multitalk.py
+++ b/wan/multitalk.py
@@ -1,6 +1,5 @@
 # Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
 import gc
-from inspect import ArgSpec
 import logging
 import json
 import math