fix(pitch): add zero-phoneme defense and logging for CUBLAS debug

- PitchInference.cpp: log phoneme count/ph_dur shape/predictor input
  shapes; return explicit InferenceInputInvalid when phoneCount == 0
- InputWord.cpp (preprocessPhonemeDurations): return error on zero
  phonemes (protects all callers: Duration/Pitch/Variance/Acoustic)
- ChainG2p TaskImpl: remove mode restriction on empty-pronunciation
  fallback — always fill missing pronunciation with original lyric
- ChainG2p-Eng config: add missing langRef=eng/default so Multig2p
  receives a valid language prefix instead of empty string

Author: wolfgitpr

domains/ds-infer/lib/util/inferutil/src/InputWord.cpp

@@ -99,6 +99,11 @@ namespace ds::infer::inferutil {
         }
 
         auto phoneCount = getPhoneCount(words);
+        if (phoneCount == 0) {
+            return srt::core::Error(srt::core::ErrorCode::InferenceInputInvalid,
+                                    "preprocessPhonemeDurations: phoneme count is zero");
+        }
+
         using TensorType = int64_t;
         auto exp = TensorHelper<TensorType>::createFor1DArray(phoneCount);
         if (!exp) {

plugins/G2P/chain/internal/V1/TaskImpl.cpp

@@ -60,8 +60,8 @@ namespace srt::g2p::plugins::ChainG2p::Internal::V1 {
             res.mode = word.mode;
             res.errorType = word.errorType;
 
-            // copy 模式下，如果发音为空，应该原样返回
-            if (res.mode == srt::g2p::kG2pModeCopy && res.pronunciation.empty()) {
+            // 无论 mode 为何，pronunciation 不应为空；兜底使用原 lyric
+            if (res.pronunciation.empty()) {
                 res.pronunciation = res.lyric;
                 if (res.candidates.empty()) {
                     res.candidates = {res.lyric};

plugins/diffsinger/pitch/PitchInference.cpp

@@ -329,10 +329,35 @@ namespace srt::svs {
             return exp.takeError();
         }
 
+        {
+            auto phoneCount = ds::infer::inferutil::getPhoneCount(pitchInput->words);
+            Log.srtInfo("%1 phoneme count = %2", kLogPrefix, phoneCount);
+            for (size_t wi = 0; wi < pitchInput->words.size(); ++wi) {
+                const auto &w = pitchInput->words[wi];
+                Log.srtInfo("%1   word[%2] phones=%3 notes=%4", kLogPrefix, wi, w.phones.size(), w.notes.size());
+                for (size_t pi = 0; pi < w.phones.size(); ++pi) {
+                    Log.srtInfo("%1     phone[%2] token='%3' lang='%4'", kLogPrefix, pi, w.phones[pi].token, w.phones[pi].language);
+                }
+            }
+            if (phoneCount == 0) {
+                setState(Failed);
+                Log.srtCritical("%1 start: zero phonemes from G2P — cannot run pitch inference", kLogPrefix);
+                return srt::core::Error::inferenceError(srt::core::ErrorCode::InferenceInputInvalid,
+                                  "[Pitch] zero phonemes from G2P — cannot run pitch inference",
+                                  {}, "pitch");
+            }
+        }
         if (auto exp = ds::infer::inferutil::preprocessPhonemeDurations(pitchInput->words,
                                                                      config->frameWidth);
             exp) {
-            sessionInput->inputs.emplace("ph_dur", exp.take());
+            auto phDur = exp.take();
+            {
+                auto shape = phDur->shape();
+                std::string shapeStr;
+                for (auto s : shape) shapeStr += std::to_string(s) + " ";
+                Log.srtInfo("%1 ph_dur shape: [%2]", kLogPrefix, shapeStr);
+            }
+            sessionInput->inputs.emplace("ph_dur", std::move(phDur));
         } else {
             setState(Failed);
             Log.srtCritical("%1 start: preprocessPhonemeDurations failed: %2",
@@ -545,6 +570,15 @@ namespace srt::svs {
                               {}, "pitch");
         }
 
+        // Log input tensor shapes before predictor inference
+        Log.srtInfo("%1 predictor input keys:", kLogPrefix);
+        for (const auto &[name, tensor] : sessionInput->inputs) {
+            auto shape = tensor->shape();
+            std::string shapeStr;
+            for (auto s : shape) shapeStr += std::to_string(s) + " ";
+            Log.srtInfo("%1   '%2' shape=[%3]", kLogPrefix, name, shapeStr);
+        }
+
         srt::core::NO<srt::core::TaskResult> sessionTaskResult;
         auto sessionExp = impl.predictorSession->start(sessionInput);
         if (!sessionExp) {

resources/G2pPackages/Phonetic-Suite-Eng/modules/ChainG2p-Eng/config.json

@@ -1 +1 @@
-{"$version": "1.0", "level": 1, "mode": "standard", "schema": {}, "configuration": {"steps": [{"step": "tagAndValidate", "params": {"tagger": [{"type": "regex", "value": ["([A-Z]+)"], "tag": "uppercase", "action": "copy"}, {"type": "regex", "value": ["([a-z]+)"], "tag": "lowercase", "action": "convert"}]}}, {"step": "dict", "params": {"enabled": true, "file": "ds_cmudict-07b.txt"}}, {"step": "model", "params": {"enabled": true, "id": "g2p-multig2p-multi-official", "batchSize": 20}}, {"step": "fallback", "params": {"useOriginal": true, "markFailed": true}}]}}
+{"$version": "1.0", "level": 1, "mode": "standard", "schema": {}, "configuration": {"steps": [{"step": "tagAndValidate", "params": {"tagger": [{"type": "regex", "value": ["([A-Z]+)"], "tag": "uppercase", "action": "copy"}, {"type": "regex", "value": ["([a-z]+)"], "tag": "lowercase", "action": "convert"}]}}, {"step": "dict", "params": {"enabled": true, "file": "ds_cmudict-07b.txt"}}, {"step": "model", "params": {"enabled": true, "id": "g2p-multig2p-multi-official", "langRef": "eng/default", "batchSize": 20}}, {"step": "fallback", "params": {"useOriginal": true, "markFailed": true}}]}}