{"id":1774,"date":"2026-06-29T12:45:11","date_gmt":"2026-06-29T04:45:11","guid":{"rendered":"https:\/\/www.izhuhn.cn\/?p=1774"},"modified":"2026-06-29T12:45:11","modified_gmt":"2026-06-29T04:45:11","slug":"minicpm-o-%e9%9b%86%e6%88%90-cosyvoice3-%e6%8e%92%e5%9d%91%e7%ae%80%e8%ae%b0","status":"publish","type":"post","link":"https:\/\/www.izhuhn.cn\/index.php\/2026\/06\/29\/minicpm-o-%e9%9b%86%e6%88%90-cosyvoice3-%e6%8e%92%e5%9d%91%e7%ae%80%e8%ae%b0\/","title":{"rendered":"MiniCPM-o \u96c6\u6210 CosyVoice3 \u6392\u5751\u7b80\u8bb0"},"content":{"rendered":"<h2 class=\"wp-block-heading\">\u4e00\u3001\u80cc\u666f<\/h2>\n\n<p class=\"wp-block-paragraph\">\u505a MiniCPM-o \u5168\u53cc\u5de5\u96c6\u6210 CosyVoice3\uff08Ascend-SACT HTTP \u540e\u7aef\uff0cRTF\u22480.27\uff09\u65f6\uff0c\u542f\u52a8 server \u76f4\u63a5\u5d29\u5728\u6a21\u578b\u52a0\u8f7d\u9636\u6bb5\u3002\u8865\u4e01\u811a\u672c\u662f\u4e00\u952e\u6253\u7684\uff0c\u4f46\u5b9e\u9645\u8dd1\u8d77\u6765\u5c31\u70b8\uff0c\u6392\u67e5\u53d1\u73b0\u8865\u4e01\u811a\u672c\u672c\u8eab\u6709\u51e0\u5904\u903b\u8f91\u7f3a\u9677\u3002<\/p>\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u62a5\u9519\u73b0\u573a<\/h2>\n\n<pre class=\"wp-block-code\"><code>AttributeError: 'DuplexCapability' object has no attribute '_cv3_client'\n\n  File \"MiniCPMO45\/modeling_minicpmo_unified.py\", line 4454, in _reset_streaming_state\n    if self._cv3_client is not None:\n       ^^^^^^^^^^^^^^^^<\/code><\/pre>\n\n<p class=\"wp-block-paragraph\">\u9519\u8bef\u53d1\u751f\u5728 <code>DuplexCapability.__init__<\/code> \u2192 <code>_reset_streaming_state()<\/code> \u8c03\u7528\u94fe\u4e2d\uff0c<code>_cv3_client<\/code> \u5c5e\u6027\u6839\u672c\u4e0d\u5b58\u5728\u3002<\/p>\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u6839\u56e0\u5206\u6790<\/h2>\n\n<p class=\"wp-block-paragraph\">\u4e00\u5171\u627e\u5230 <strong>4 \u4e2a bug<\/strong>\uff0c\u5176\u4e2d 3 \u4e2a\u662f\u8865\u4e01\u811a\u672c <code>apply_cosyvoice3_patch.py<\/code> \u7684\u9519\u8bef\u5bfc\u81f4\uff0c1 \u4e2a\u662f\u8fd0\u884c\u65f6\u903b\u8f91\u7f3a\u9677\u3002<\/p>\n\n<h3 class=\"wp-block-heading\">Bug 1\uff1aCV3 \u521d\u59cb\u5316\u653e\u9519\u4e86\u65b9\u6cd5\uff08\u81f4\u547d\uff09<\/h3>\n\n<p class=\"wp-block-paragraph\">\u8865\u4e01\u811a\u672c\u610f\u56fe\u628a CV3 \u5ba2\u6237\u7aef\u521d\u59cb\u5316\u63d2\u5165 <code>__init__<\/code>\uff0c\u4f46\u641c\u7d22\u903b\u8f91\u6709 bug\uff1a<\/p>\n\n<pre class=\"wp-block-code\"><code>for i, line in enumerate(lines):\n    if \"self._reset_streaming_state()\" in line:\n        insert_after = i  # \u4e0d\u65ad\u8986\u76d6\uff0c\u6700\u7ec8\u62ff\u5230\u6700\u540e\u4e00\u4e2a\u5339\u914d<\/code><\/pre>\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u5faa\u73af\u904d\u5386\u6574\u4e2a\u6587\u4ef6\u627e <code>_reset_streaming_state()<\/code> \u8c03\u7528\uff0c\u4f46 <code>__init__<\/code> \u548c <code>prepare()<\/code> \u4e24\u4e2a\u65b9\u6cd5\u90fd\u6709\u8fd9\u4e2a\u8c03\u7528\u3002\u6700\u540e\u4e00\u6b21\u8986\u76d6\u53d1\u751f\u5728 <code>prepare()<\/code> \u91cc\uff0c\u7ed3\u679c CV3 \u521d\u59cb\u5316\u88ab\u9519\u8bef\u5730\u63d2\u5165\u5230\u4e86 <code>prepare()<\/code> \u65b9\u6cd5\u4e2d\u3002<\/p>\n\n<p class=\"wp-block-paragraph\">\u540e\u679c\uff1a<\/p>\n\n<ul class=\"wp-block-list\"><li><code>__init__<\/code> \u8c03\u7528 <code>_reset_streaming_state()<\/code> \u65f6\uff0c<code>_cv3_client<\/code> \u4ece\u672a\u88ab\u5b9a\u4e49 \u2192 <strong>AttributeError \u76f4\u63a5\u5d29<\/strong><\/li><li><code>prepare()<\/code> \u7b7e\u540d\u4e3a\u56fa\u5b9a\u53c2\u6570\uff08\u6ca1\u6709 <code>**kwargs<\/code>\uff09\uff0c\u91cc\u9762\u7684 <code>kwargs.get(...)<\/code> \u5373\u4f7f\u6267\u884c\u5230\u4e5f\u4f1a <code>NameError<\/code><\/li><\/ul>\n\n<h3 class=\"wp-block-heading\">Bug 2\uff1astreaming_generate \u7684 TTS \u7ba1\u9053\u6ca1\u88ab\u66ff\u6362<\/h3>\n\n<p class=\"wp-block-paragraph\">\u8865\u4e01\u811a\u672c\u7684 <code>patch_streaming_generate()<\/code> \u6709\"\u5df2\u6253\u8fc7\u8865\u4e01\"\u68c0\u6d4b\uff1a<\/p>\n\n<pre class=\"wp-block-code\"><code>if \"self._cv3_client is not None\" in content:\n    print(\"\u2705 streaming_generate() \u5df2\u6253\u8fc7\u8865\u4e01\uff0c\u8df3\u8fc7\")\n    return True<\/code><\/pre>\n\n<p class=\"wp-block-paragraph\">\u4f46\u662f <code>_reset_streaming_state<\/code> \u548c <code>prepare<\/code> \u4e24\u4e2a\u65b9\u6cd5\u91cc\u4e5f\u5305\u542b <code>self._cv3_client is not None<\/code>\u3002\u5b83\u4eec\u7684\u8865\u4e01\u5148\u6267\u884c\uff0c\u7ed3\u679c <code>streaming_generate<\/code> \u7684\u68c0\u6d4b\u88ab<strong>\u5047\u9633\u6027<\/strong>\u6b3a\u9a97\uff0c\u76f4\u63a5\u8df3\u8fc7\u4e86\u3002<\/p>\n\n<p class=\"wp-block-paragraph\">\u540e\u679c\uff1a<code>streaming_generate()<\/code> \u91cc\u7684 TTS \u7ba1\u9053\u6839\u672c\u6ca1\u6709\u88ab\u66ff\u6362\uff0cCV3 \u5ba2\u6237\u7aef\u521b\u5efa\u4e86\u4f46\u4ece\u672a\u88ab\u8c03\u7528\u3002\u5373\u4f7f\u8fc7\u4e86 Bug 1\uff0c\u8fd0\u884c\u65f6\u4e5f\u53ea\u4f1a\u8d70\u539f\u59cb MiniCPMTTS + Token2Wav \u7ba1\u9053\uff0cCosyVoice3 \u5f62\u540c\u865a\u8bbe\u3002<\/p>\n\n<h3 class=\"wp-block-heading\">Bug 3\uff1aElse \u7f29\u8fdb\u7834\u574f\uff08\u8865\u4e01\u4ee3\u7801\u672c\u8eab\u7684\u7ed3\u6784\u95ee\u9898\uff09<\/h3>\n\n<p class=\"wp-block-paragraph\">\u5373\u4f7f Bug 2 \u7684\u68c0\u6d4b\u88ab\u4fee\u6b63\uff0c\u8865\u4e01\u63d2\u5165\u7684\u4ee3\u7801\u7ed3\u6784\u672c\u8eab\u4e5f\u6709\u95ee\u9898\uff1a<\/p>\n\n<pre class=\"wp-block-code\"><code>if self._cv3_client is not None:\n    # ... CV3 pipeline ...\n    return result\n\n# \u539f\u59cb MiniCPMTTS + Token2Wav \u7ba1\u9053\uff08\u4fdd\u7559\u4f5c\u4e3a fallback\uff09\nelse:\n    # TTS generate  \u2190 \u8fd9\u91cc\u7684\u539f\u59cb\u4ee3\u7801\u6ca1\u6709\u7f29\u8fdb\u5230 else \u5757\u5185!<\/code><\/pre>\n\n<p class=\"wp-block-paragraph\">\u8865\u4e01\u53ea\u63d2\u5165\u4e86 <code>else:<\/code> \u4f46<strong>\u6ca1\u6709\u589e\u52a0\u539f\u59cb\u4ee3\u7801\u7684\u7f29\u8fdb\u5c42\u7ea7<\/strong>\u3002\u7ed3\u679c <code>else<\/code> \u5757\u662f\u7a7a\u7684\uff0c\u539f\u59cb TTS \u4ee3\u7801\u65e0\u6761\u4ef6\u6267\u884c\u3002<\/p>\n\n<p class=\"wp-block-paragraph\">\u4fee\u590d\u65b9\u6848\uff1a\u4e0d\u7528 <code>else<\/code>\uff0c\u6539\u7528 <strong>early return<\/strong> \u6a21\u5f0f\u2014\u2014CV3 \u8def\u5f84\u8d70\u5b8c\u76f4\u63a5 return\uff0c\u539f\u59cb\u4ee3\u7801\u539f\u6837\u4fdd\u7559\u4f5c\u4e3a fallback\uff0c\u4e0d\u7834\u574f\u4efb\u4f55\u7f29\u8fdb\u3002<\/p>\n\n<h3 class=\"wp-block-heading\">Bug 4\uff1astreaming_prefill \u9519\u8bef\u91cd\u7f6e\u6587\u672c buffer<\/h3>\n\n<p class=\"wp-block-paragraph\">CV3 \u7684\u6587\u672c\u7d2f\u79ef\u673a\u5236\u662f\u8de8 unit \u7684\uff1a\u6bcf\u4e2a <code>streaming_generate<\/code> chunk \u7684\u6587\u672c\u901a\u8fc7 <code>feed_text()<\/code> \u8ffd\u52a0\u5230 buffer\uff0c\u6512\u591f\u5b57\u7b26\u6570\u540e\u89e6\u53d1\u5408\u6210\u3002<\/p>\n\n<p class=\"wp-block-paragraph\">\u4f46\u6bcf\u6b21 <code>streaming_prefill<\/code> \u8c03\u7528\u65f6\u4f1a\u628a <code>_text_buffer<\/code> \u6e05\u7a7a\uff1a<\/p>\n\n<pre class=\"wp-block-code\"><code># streaming_prefill() \u4e2d\uff1a\nif self._cv3_client is not None:\n    self._cv3_client._text_buffer = \"\"  # \u2190 \u6e05\u7a7a\u4e86\u8de8 unit \u7d2f\u79ef\u7684\u6587\u672c!<\/code><\/pre>\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u5355\u6b21 <code>streaming_generate<\/code> \u751f\u6210\u7684\u6587\u672c\u957f\u5ea6\u5c0f\u4e8e <code>min_text_len_for_synth<\/code>\uff08\u9ed8\u8ba4 2 \u5b57\u7b26\uff09\uff0c\u6587\u672c\u88ab\u5b58\u5165 buffer\u3002\u4e0b\u4e00\u4e2a unit \u7684 <code>streaming_prefill<\/code> \u76f4\u63a5\u628a\u5b83\u6e05\u4e86\uff0c\u6587\u672c\u6c38\u8fdc\u6512\u4e0d\u591f\uff0c<strong>TTS \u5408\u6210\u6c38\u8fdc\u4e0d\u4f1a\u89e6\u53d1<\/strong>\u3002<\/p>\n\n<p class=\"wp-block-paragraph\">\u4fee\u590d\uff1a\u6587\u672c buffer \u53ea\u5728 <code>_reset_streaming_state()<\/code>\uff08session \u91cd\u542f\u65f6\uff09\u548c <code>flush_text()<\/code>\uff08\u5408\u6210\u65f6\uff09\u6e05\u7a7a\u3002<\/p>\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u4fee\u590d\u6e05\u5355<\/h2>\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>\u6587\u4ef6<\/th><th>\u4fee\u590d\u5185\u5bb9<\/th><\/tr><\/thead>\n<tbody>\n<tr><td><code>modeling_minicpmo_unified.py<\/code><\/td><td>CV3 \u521d\u59cb\u5316\u4ece <code>prepare()<\/code> \u79fb\u5230 <code>__init__()<\/code>\uff0c\u653e\u5728 <code>_reset_streaming_state()<\/code> \u4e4b\u524d<\/td><\/tr>\n<tr><td><code>modeling_minicpmo_unified.py<\/code><\/td><td>\u5728 <code>streaming_generate()<\/code> \u7684 <code># TTS generate<\/code> \u4e4b\u524d\u63d2\u5165 CV3 TTS \u7ba1\u9053\uff08early return \u6a21\u5f0f\uff09<\/td><\/tr>\n<tr><td><code>modeling_minicpmo_unified.py<\/code><\/td><td>\u79fb\u9664 <code>streaming_prefill()<\/code> \u4e2d\u7684 <code>_text_buffer<\/code> \u91cd\u7f6e<\/td><\/tr>\n<tr><td><code>apply_cosyvoice3_patch.py<\/code><\/td><td>\u4fee\u590d <code>patch_duplex_capability_init()<\/code>\uff1a\u9650\u5b9a\u5728 <code>__init__<\/code> \u65b9\u6cd5\u5185\u641c\u7d22\uff0c\u63d2\u5165\u5728 <code>_reset_streaming_state()<\/code> \u4e4b\u524d<\/td><\/tr>\n<tr><td><code>apply_cosyvoice3_patch.py<\/code><\/td><td>\u4fee\u590d <code>patch_streaming_generate()<\/code>\uff1a\u7cbe\u786e\u68c0\u6d4b\u5df2\u6253\u8865\u4e01\uff08\u68c0\u67e5\u90bb\u8fd1\u884c\uff09\uff0c\u6539\u7528 early return \u66ff\u4ee3 else \u7f29\u8fdb<\/td><\/tr>\n<tr><td><code>apply_cosyvoice3_patch.py<\/code><\/td><td>\u4fee\u590d f-string \u8f6c\u4e49\u8bed\u6cd5\u9519\u8bef<\/td><\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u6700\u7ec8\u6587\u4ef6\u7ed3\u6784<\/h2>\n\n<pre class=\"wp-block-code\"><code>DuplexCapability.__init__:\n  self._use_cosyvoice3 = kwargs.get(\"use_cosyvoice3\", False)\n  self._cv3_client = None\n  if self._use_cosyvoice3:   # \u2190 \u5148\u521d\u59cb\u5316\n      self._cv3_client = DuplexCosyVoice3TTS(...)\n  self._reset_streaming_state()  # \u2190 \u518d\u91cd\u7f6e\uff0c\u6b64\u65f6 _cv3_client \u5df2\u5b9a\u4e49\n\nstreaming_generate:\n  if self._cv3_client is not None:    # \u2190 CV3 \u7ba1\u9053\uff0cearly return\n      clean_text = strip_tts_tokens(text)\n      self._cv3_client.feed_text(clean_text)\n      if self._cv3_client.should_synthesize(force=end_of_turn):\n          audio = self._cv3_client.synthesize_turn(...)\n      return result\n  # TTS generate \u2190 \u539f\u59cb\u7ba1\u9053 fallback\uff0c\u7f29\u8fdb\u4e0d\u53d8<\/code><\/pre>\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u6559\u8bad<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>\u8865\u4e01\u811a\u672c\u7684\"\u5df2\u6253\u8865\u4e01\"\u68c0\u6d4b\u4e0d\u80fd\u9760\u5168\u5c40\u5173\u952e\u5b57\u5339\u914d<\/strong>\u2014\u2014\u540c\u4e00\u4e2a\u5b57\u7b26\u4e32\u53ef\u80fd\u51fa\u73b0\u5728\u5176\u4ed6\u5df2\u6253\u8865\u4e01\u7684\u4f4d\u7f6e\uff0c\u5bfc\u81f4\u5047\u9633\u6027\u3002<\/li><li><strong>\u63d2\u5165\u4ee3\u7801\u8981\u6ce8\u610f\u7f29\u8fdb\u7834\u574f<\/strong>\u2014\u2014\u5f80\u5df2\u6709\u4ee3\u7801\u524d\u63d2\u5165 if\/else \u65f6\uff0c\u539f\u6709\u4ee3\u7801\u7684\u7f29\u8fdb\u4e0d\u4f1a\u81ea\u52a8\u53d8\u66f4\u3002\u7528 early return \u6a21\u5f0f\u6bd4 else \u5305\u88f9\u66f4\u5b89\u5168\u3002<\/li><li><strong>\u6587\u4ef6\u7ea7\u522b\u7684 for \u5faa\u73af\u627e\"\u6700\u540e\u4e00\u4e2a\"\u5339\u914d\u5f88\u5371\u9669<\/strong>\u2014\u2014\u540c\u540d\u65b9\u6cd5\u8c03\u7528\u53ef\u80fd\u5728\u5176\u4ed6\u65b9\u6cd5\u4e2d\u51fa\u73b0\uff0c\u5e94\u8be5\u5148\u5b9a\u4f4d\u65b9\u6cd5\u8fb9\u754c\u518d\u641c\u7d22\u3002<\/li><\/ul>\n\n<p class=\"wp-block-paragraph\">\u7b80\u8bb0\u3002<\/p>","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u80cc\u666f \u505a MiniCPM-o \u5168\u53cc\u5de5\u96c6\u6210 CosyVoice3\uff08Ascend-SACT HTTP \u540e\u7aef\uff0cRTF\u22480.27\uff09\u65f6\uff0c &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[1],"tags":[40,42,39,15,41],"class_list":["post-1774","post","type-post","status-publish","format-standard","hentry","category-uncategorized","tag-cosyvoice3","tag-debug","tag-minicpm","tag-npu","tag-tts"],"jetpack_featured_media_url":"","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1774","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1774"}],"version-history":[{"count":1,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1774\/revisions"}],"predecessor-version":[{"id":1775,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1774\/revisions\/1775"}],"wp:attachment":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1774"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1774"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1774"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}