Kazeia
/
kazeia
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Reduce talker KV_LEN 100→64: saves 148ms (RTF 1.31) 

KV window of 64 sufficient for ~70 token generation (10 prefill + 58 gen).
36% less KV memcpy per talker step (28L × 2 × 64×8×128 vs 100×8×128).

Generation: 3795ms → 3647ms, total: 6438ms → 6093ms

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Kazeia Team 2026-04-09 12:47:30 +02:00
parent 4dcc4bb8b3
commit a688edc9ec
3 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
executorch-custom/jni_layer_tts.cpp
View File

@ -692,7 +692,7 @@ ExecuTorchJni::runTtsPipelineImpl(
jint maxTokens) jint maxTokens)
{ {
static const int DIM=1024,VOCAB=3072,CB_SIZE=2048,NUM_CB=16; static const int DIM=1024,VOCAB=3072,CB_SIZE=2048,NUM_CB=16;
static const int T_L=28,T_KV=8,T_HD=128,T_KV_LEN=100; static const int T_L=28,T_KV=8,T_HD=128,T_KV_LEN=64;
static const int C_L=5,C_KV=8,C_HD=128,C_KV_LEN=16; static const int C_L=5,C_KV=8,C_HD=128,C_KV_LEN=16;
static const int CODEC_EOS=2150; static const int CODEC_EOS=2150;

2
kazeia-android/app/src/main/java/com/kazeia/tts/Qwen3TtsEngine.kt
View File

@ -68,7 +68,7 @@ class Qwen3TtsEngine(
private const val CP_KV_LEN = 16 // max 16 past positions (17 total with current) private const val CP_KV_LEN = 16 // max 16 past positions (17 total with current)
// Talker .pte constants // Talker .pte constants
private const val TALKER_PTE_KV_LEN = 100 // .pte talker KV window size private const val TALKER_PTE_KV_LEN = 64 // .pte talker KV window size (reduced from 100)
// Codec special token IDs (in talker's 3072 vocab space) // Codec special token IDs (in talker's 3072 vocab space)
private const val CODEC_EOS = 2150 private const val CODEC_EOS = 2150

4
scripts/export_talker_pte.py
View File

@ -11,7 +11,7 @@ warnings.filterwarnings('ignore')
N_L = 28; N_H = 16; N_KV = 8; HD = 128; DIM = 1024; N_REP = 2 N_L = 28; N_H = 16; N_KV = 8; HD = 128; DIM = 1024; N_REP = 2
VOCAB = 3072; FFN = 3072 VOCAB = 3072; FFN = 3072
KV_LEN = 16 # Small KV for testing HTP viability KV_LEN = 64 # Reduced from 100: saves 36% memcpy, sufficient for ~70 token generation
state = torch.load("/opt/Kazeia/models_qnn/qwen3-tts-export/qwen3_tts_talker.pth", state = torch.load("/opt/Kazeia/models_qnn/qwen3-tts-export/qwen3_tts_talker.pth",
map_location="cpu", weights_only=False) map_location="cpu", weights_only=False)
@ -117,7 +117,7 @@ edge = to_edge_transform_and_lower_to_qnn(w, (e, m, c0, s0, *kvs), compiler_spec
print("LOWERED!") print("LOWERED!")
pte = edge.to_executorch() pte = edge.to_executorch()
OUT = "/opt/Kazeia/models_qnn/talker_transformer_fp16_kv16.pte" OUT = "/opt/Kazeia/models_qnn/talker_transformer_fp16.pte"
with open(OUT, "wb") as f: with open(OUT, "wb") as f:
pte.write_to_file(f) pte.write_to_file(f)
print(f"SAVED: {OUT} ({os.path.getsize(OUT)/1024/1024:.0f} MB)") print(f"SAVED: {OUT} ({os.path.getsize(OUT)/1024/1024:.0f} MB)")