Narrate Both Sides
diff --git a/koboldcpp.py b/koboldcpp.py
index e70a28139..b79c72649 100644
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -783,7 +783,6 @@ def fetch_gpu_properties(testCL,testCU,testVK):
FetchedCUdeviceMem = [line.split(",")[1].strip().split(" ")[0].strip() for line in output.splitlines()]
FetchedCUfreeMem = [line.split(",")[2].strip().split(" ")[0].strip() for line in output.splitlines()]
except Exception:
- FetchedCUdevices = []
FetchedCUdeviceMem = []
FetchedCUfreeMem = []
faileddetectvram = True
@@ -806,7 +805,6 @@ def fetch_gpu_properties(testCL,testCU,testVK):
if getamdvram:
FetchedCUdeviceMem = [line.split(",")[1].strip() for line in getamdvram.splitlines()[1:] if line.strip()]
except Exception:
- FetchedCUdevices = []
FetchedCUdeviceMem = []
FetchedCUfreeMem = []
faileddetectvram = True
@@ -817,6 +815,8 @@ def fetch_gpu_properties(testCL,testCU,testVK):
for idx in range(0,4):
if(len(FetchedCUdevices)>idx):
CUDevicesNames[idx] = FetchedCUdevices[idx]
+ for idx in range(0,4):
+ if(len(FetchedCUdevices)>idx):
if len(FetchedCUdeviceMem)>idx:
dmem = int(FetchedCUdeviceMem[idx]) if AMDgpu else (int(FetchedCUdeviceMem[idx])*1024*1024)
lowestcumem = dmem if lowestcumem==0 else (dmem if dmem
llama_v2_tokenize(struct llama_v2_context * ctx, const
res.resize(n);
return res;
-}
+}
\ No newline at end of file
diff --git a/otherarch/llama_v3.cpp b/otherarch/llama_v3.cpp
index d0dd2f994..b1e976afc 100644
--- a/otherarch/llama_v3.cpp
+++ b/otherarch/llama_v3.cpp
@@ -4414,3 +4414,16 @@ static void llama_v3_log_callback_default(llama_v3_log_level level, const char *
fputs(text, stderr);
fflush(stderr);
}
+
+//// stuff this here since it's just some obsolete junk ////
+static std::vector kcpp_compute_buf;
+void kcpp_graph_compute_helper(struct ggml_v3_cgraph *graph, int n_threads)
+{
+ struct ggml_v3_cplan plan = ggml_v3_graph_plan(graph, n_threads);
+ if (plan.work_size > 0)
+ {
+ kcpp_compute_buf.resize(plan.work_size);
+ plan.work_data = kcpp_compute_buf.data();
+ }
+ ggml_v3_graph_compute(graph, &plan);
+}
\ No newline at end of file
diff --git a/otherarch/tts_adapter.cpp b/otherarch/tts_adapter.cpp
index 4311346ac..7db7dbffb 100644
--- a/otherarch/tts_adapter.cpp
+++ b/otherarch/tts_adapter.cpp
@@ -574,6 +574,10 @@ tts_generation_outputs ttstype_generate(const tts_generation_inputs inputs)
}
}
+ double ttstime = 0;
+ timer_start();
+
+
if(!inputs.quiet && ttsdebugmode==1)
{
printf("\nInput: %s\n", prompt_clean.c_str());
@@ -591,6 +595,14 @@ tts_generation_outputs ttstype_generate(const tts_generation_inputs inputs)
{
printf("\nReuse speaker ID=%d (%d tokens)...", last_speaker_seed, last_speaker_codes.size());
}
+ } else if (speaker_seed==1){ //1 is a special seed
+ std::string speaker = "but<|t_0.31|><|code_start|><|1023|><|1474|><|17|><|121|><|1362|><|744|><|438|><|1319|><|744|><|1419|><|1246|><|923|><|1338|><|406|><|939|><|975|><|1491|><|965|><|1212|><|248|><|794|><|464|><|830|><|code_end|>\nthat<|t_0.13|><|code_start|><|1578|><|1773|><|660|><|1074|><|221|><|1803|><|142|><|914|><|798|><|485|><|code_end|>\nis<|t_0.11|><|code_start|><|737|><|794|><|1288|><|182|><|895|><|1653|><|448|><|471|><|code_end|>\nwhat<|t_0.12|><|code_start|><|1734|><|1306|><|779|><|490|><|525|><|1028|><|37|><|1633|><|1353|><|code_end|>\nit<|t_0.09|><|code_start|><|1343|><|898|><|270|><|1035|><|94|><|1409|><|388|><|code_end|>\nis<|t_0.23|><|code_start|><|694|><|695|><|577|><|692|><|1047|><|388|><|28|><|905|><|1155|><|50|><|1629|><|1775|><|1711|><|1729|><|404|><|1027|><|344|><|code_end|>";
+ last_speaker_codes = common_tokenize(model_ttc, speaker, false, true);
+ last_speaker_seed = speaker_seed;
+ if(!inputs.quiet && ttsdebugmode==1)
+ {
+ printf("\nSpecial ID=%d (%d tokens)...", last_speaker_seed, last_speaker_codes.size());
+ }
} else {
//generate the voice texture of our new speaker
last_speaker_codes.clear();
@@ -800,8 +812,8 @@ tts_generation_outputs ttstype_generate(const tts_generation_inputs inputs)
const int n_sr = 24000; // sampling rate
- // zero out first 0.25 seconds or 0.05 depending on whether its seeded
- const int cutout = (speaker_seed>0?(24000/4):(24000/20));
+ // zero out first 0.2 seconds or 0.05 depending on whether its seeded
+ const int cutout = (speaker_seed>0?(24000/5):(24000/20));
for (int i = 0; i < cutout; ++i) {
audio[i] = 0.0f;
}
@@ -811,10 +823,11 @@ tts_generation_outputs ttstype_generate(const tts_generation_inputs inputs)
}
last_generated_audio = save_wav16_base64(audio, n_sr);
+ ttstime = timer_check();
if(!inputs.quiet)
{
- printf("\nTTS Generated %d audio tokens.\n",(int) codes.size());
+ printf("\nTTS Generated %d audio tokens in %.2fs.\n",(int) codes.size(),ttstime);
}
output.data = last_generated_audio.c_str();
diff --git a/otherarch/utils.cpp b/otherarch/utils.cpp
index f848eeabd..12296360e 100644
--- a/otherarch/utils.cpp
+++ b/otherarch/utils.cpp
@@ -225,18 +225,6 @@ bool should_transpose_layer(std::string name)
return false;
}
-static std::vector kcpp_compute_buf;
-void kcpp_graph_compute_helper(struct ggml_v3_cgraph *graph, int n_threads)
-{
- struct ggml_v3_cplan plan = ggml_v3_graph_plan(graph, n_threads);
- if (plan.work_size > 0)
- {
- kcpp_compute_buf.resize(plan.work_size);
- plan.work_data = kcpp_compute_buf.data();
- }
- ggml_v3_graph_compute(graph, &plan);
-}
-
static const std::string kcpp_base64_chars =
"ABCDEFGHIJKLMNOPQRSTUVWXYZ"
"abcdefghijklmnopqrstuvwxyz"