Cherry-pick for 1.17.1 patch release (microsoft#19477)

### Description  As title. ### Motivation and Context  --------- Co-authored-by: petermcaughan <[email protected]> Co-authored-by: Peter McAughan <[email protected]> Co-authored-by: Adrian Lizarraga <[email protected]> Co-authored-by: Patrice Vignola <[email protected]> Co-authored-by: ivberg <[email protected]> Co-authored-by: Yulong Wang <[email protected]> Co-authored-by: Baiju Meswani <[email protected]> Co-authored-by: Preetha Veeramalai <[email protected]> Co-authored-by: fxmarty <[email protected]> Co-authored-by: Sheil Kumar <[email protected]> Co-authored-by: Sheil Kumar <[email protected]> Co-authored-by: Prathik Rao <[email protected]> Co-authored-by: Shubham Bhokare <[email protected]> Co-authored-by: kunal-vaishnavi <[email protected]> Co-authored-by: Jian Chen <[email protected]> Co-authored-by: Xavier Dupré <[email protected]> Co-authored-by: satyajandhyala <[email protected]>
mo-ja · Feb 21, 2024 · 75968b9 · 75968b9
1 parent 5f0b62c
commit 75968b9
Show file tree

Hide file tree

Showing 76 changed files with 2,220 additions and 579 deletions.
diff --git a/cmake/adjust_global_compile_flags.cmake b/cmake/adjust_global_compile_flags.cmake
@@ -92,8 +92,13 @@ if (onnxruntime_MINIMAL_BUILD)
   endif()
 endif()
 
-# enable stream for all the non-minimal build
-if (NOT onnxruntime_MINIMAL_BUILD)
+# Enable stream for all the non-minimal build, except for DML. There's currently a bug
+# in the allocation planner when reusing buffers and more than one streams are used that
+# make it possible (although rarely) to reach a reference count of 0 for a buffer that is
+# still being used. Since DML doesn't benefit from multiple streams, disabling it is the
+# safest option for now.
+# https://github.com/microsoft/onnxruntime/issues/19480
+if (NOT onnxruntime_MINIMAL_BUILD AND NOT onnxruntime_USE_DML)
   add_compile_definitions(ORT_ENABLE_STREAM)
 endif()
 

diff --git a/docs/ContribOperators.md b/docs/ContribOperators.md
@@ -461,7 +461,7 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dt><tt>repetition_penalty</tt> (optional) : T</dt>
 <dd>The parameter for repetition penalty. Default value 1.0 means no penalty. Accepts value > 0.0. Shape is (1)</dd>
 <dt><tt>vocab_mask</tt> (optional) : M</dt>
-<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vacab_size)</dd>
+<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vocab_size)</dd>
 <dt><tt>prefix_vocab_mask</tt> (optional) : M</dt>
 <dd>Mask of vocabulary for first step. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (batch_size, vocab_size)</dd>
 <dt><tt>attention_mask</tt> (optional) : I</dt>
@@ -2252,7 +2252,7 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dt><tt>repetition_penalty</tt> (optional) : T</dt>
 <dd>The parameter for repetition penalty. Default value 1.0 means no penalty. Accepts value > 0.0. Shape is (1)</dd>
 <dt><tt>vocab_mask</tt> (optional) : I</dt>
-<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vacab_size)</dd>
+<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vocab_size)</dd>
 <dt><tt>prefix_vocab_mask</tt> (optional) : I</dt>
 <dd>Mask of vocabulary for first step. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (batch_size, vocab_size)</dd>
 <dt><tt>attention_mask</tt> (optional) : I</dt>
@@ -5154,7 +5154,7 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dt><tt>repetition_penalty</tt> (optional) : T</dt>
 <dd>The parameter for repetition penalty. Default value 1.0 means no penalty. Accepts value > 0.0. Shape is (1)</dd>
 <dt><tt>vocab_mask</tt> (optional) : I</dt>
-<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vacab_size)</dd>
+<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vocab_size)</dd>
 <dt><tt>prefix_vocab_mask</tt> (optional) : I</dt>
 <dd>Mask of vocabulary for first step. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (batch_size, vocab_size)</dd>
 <dt><tt>attention_mask</tt> (optional) : I</dt>
@@ -5743,12 +5743,14 @@ This version of the operator has been available since version 1 of the 'com.micr
 #### Attributes
 
 <dl>
+<dt><tt>beginning_timestamp_token_id</tt> : int</dt>
+<dd>The id of the first timestamp</dd>
 <dt><tt>decoder</tt> : graph (required)</dt>
 <dd>Decoder subgraph to execute in a loop.</dd>
 <dt><tt>decoder_output_cross_qk</tt> : int</dt>
 <dd>If nozero, decoder subgraph contains output Q*K from cross attentions. Default 0.</dd>
 <dt><tt>decoder_start_token_id</tt> : int</dt>
-<dd>The id of the token that indicates decoding starts.</dd>
+<dd>The id of the token that indicates decoding starts (i.e. the start of transcription token id)</dd>
 <dt><tt>early_stopping</tt> : int</dt>
 <dd>early stop or not</dd>
 <dt><tt>encoder</tt> : graph</dt>
@@ -5761,10 +5763,18 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dd>Must be 2 for whisper</dd>
 <dt><tt>no_repeat_ngram_size</tt> : int</dt>
 <dd>no repeat ngrams size</dd>
-<dt><tt>no_speech_token</tt> : int</dt>
+<dt><tt>no_speech_token_id</tt> : int</dt>
 <dd>The token in whisper model that marks all sequence empty. With this model, whisper could output no_speech_prob after. Default -1.</dd>
+<dt><tt>no_timestamps_token_id</tt> : int</dt>
+<dd>The id of the token that indicates no timestamps</dd>
 <dt><tt>pad_token_id</tt> : int (required)</dt>
 <dd>The id of the padding token</dd>
+<dt><tt>start_of_lm_token_id</tt> : int</dt>
+<dd>The id of the token that indicates LM starts</dd>
+<dt><tt>transcribe_token_id</tt> : int</dt>
+<dd>The id of the transcribe task</dd>
+<dt><tt>translate_token_id</tt> : int</dt>
+<dd>The id of the translate task</dd>
 <dt><tt>vocab_size</tt> : int</dt>
 <dd>Size of the vocabulary. If not provided, it will be inferred from the decoder subgraph's output shape</dd>
 </dl>
@@ -5783,11 +5793,11 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dt><tt>num_return_sequences</tt> : I</dt>
 <dd>The number of returned sequences in the batch. Shape is (1)</dd>
 <dt><tt>length_penalty</tt> (optional) : T</dt>
-<dd>Exponential penalty to the length. Default value 1.0 means no penalty.Value > 1.0 encourages longer sequences, while values < 1.0 produces shorter sequences.Shape is (1,)</dd>
+<dd>Exponential penalty to the length. Default value 1.0 means no penalty. Value > 1.0 encourages longer sequences, while values < 1.0 produces shorter sequences. Shape is (1,)</dd>
 <dt><tt>repetition_penalty</tt> (optional) : T</dt>
 <dd>The parameter for repetition penalty. Default value 1.0 means no penalty. Accepts value > 0.0. Shape is (1)</dd>
 <dt><tt>vocab_mask</tt> (optional) : M</dt>
-<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vacab_size)</dd>
+<dd>Mask of vocabulary. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (vocab_size)</dd>
 <dt><tt>prefix_vocab_mask</tt> (optional) : M</dt>
 <dd>Mask of vocabulary for first step. Words that masked with 0 are not allowed to be generated, and 1 is allowed. Shape is (batch_size, vocab_size)</dd>
 <dt><tt>attention_mask</tt> (optional) : I</dt>
@@ -5797,7 +5807,7 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dt><tt>logits_processor</tt> (optional) : I</dt>
 <dd>Specific logits processor for different types of beamsearch models. Default value 0 means no specific logit processor. Accepts value >= 0. Shape is (1)</dd>
 <dt><tt>cross_qk_layer_head</tt> (optional) : I</dt>
-<dd>Only keep this list of (layer, head) of QK in the final cross_qk output when use_cross_qk is set. Default collect allits shape is (number of (layer, head) to keep, 2), i.e., [[layer_id1, head_id1], [layer_id2, head_id2]......]</dd>
+<dd>Only keep this list of (layer, head) of QK in the final cross_qk output when use_cross_qk is set. Default collect all its shape is (number of (layer, head) to keep, 2), i.e., [[layer_id1, head_id1], [layer_id2, head_id2]......]</dd>
 <dt><tt>extra_decoding_ids</tt> (optional) : I</dt>
 <dd>Part of the decoder_input_ids that we need cross qk for it. it is of shape  (batch_size, extra_decoding_ids_len).In such case, we should remove this from the tail of the decoder_input_ids, and put it here. ids < 0 in it (for multiple batch) are treated as stop of the extra_decoding_ids for corresponding batch.</dd>
 </dl>
@@ -5810,11 +5820,11 @@ This version of the operator has been available since version 1 of the 'com.micr
 <dt><tt>sequences_scores</tt> (optional) : T</dt>
 <dd>Final beam score of the generated sequences. Shape is (batch_size, num_return_sequences)</dd>
 <dt><tt>scores</tt> (optional) : T</dt>
-<dd>Processed beam scores for each vocabulary token at each generation step.Beam scores consisting of log softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam.Shape is (max_length - sequence_length, batch_size, num_beams, vocab_size)</dd>
+<dd>Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam. Shape is (max_length - sequence_length, batch_size, num_beams, vocab_size)</dd>
 <dt><tt>cross_qk</tt> (optional) : V</dt>
-<dd>Output the accumulated stacked Q*K in cross attentions. Let H = number of Head of cross attention, F = the frames or kv-seq-len of the cross attention input, T = real decoded token length, L = number of layers,B = batch size, R = num_return_sequences. It then should return tensor of shape [B, R, L*H, T, F].If cross_qk_layer_head is given, shape is [B, R, cross_qk_layer_head.shape[0], T, F]</dd>
+<dd>Output the accumulated stacked Q*K in cross attentions. Let H = number of Head of cross attention, F = the frames or kv-seq-len of the cross attention input, T = real decoded token length, L = number of layers, B = batch size, R = num_return_sequences. It then should return tensor of shape [B, R, L*H, T, F]. If cross_qk_layer_head is given, shape is [B, R, cross_qk_layer_head.shape[0], T, F]</dd>
 <dt><tt>non_speech_probs</tt> (optional) : T</dt>
-<dd>For whisper model, output the probabilities from logits after encoder and context decoding for the no_speech_token.Currently we treat the last token's logits is what we need, in future extra graph logic may be add to the encoder/context-decoder subgraph.The prob is save before logits may be updated by extra-decoding-ids. The shape of non_speech_probs is [B]</dd>
+<dd>For whisper model, output the probabilities from logits after encoder and context decoding for the no_speech_token_id. The shape of non_speech_probs is [B]</dd>
 </dl>
 
 #### Type Constraints

diff --git a/js/web/package.json b/js/web/package.json
@@ -69,46 +69,56 @@
   "exports": {
     ".": {
       "node": "./dist/ort.node.min.js",
+      "types": "./types.d.ts",
       "default": {
         "import": "./dist/esm/ort.min.js",
         "require": "./dist/cjs/ort.min.js",
+        "types": "./types.d.ts",
         "default": {
           "development": "./dist/ort.js",
+          "types": "./types.d.ts",
           "default": "./dist/ort.min.js"
         }
       }
     },
     "./experimental": {
       "import": "./dist/esm/ort.all.min.js",
       "require": "./dist/cjs/ort.all.min.js",
+      "types": "./types.d.ts",
       "default": {
         "development": "./dist/ort.all.js",
+        "types": "./types.d.ts",
         "default": "./dist/ort.all.min.js"
       }
     },
     "./wasm": {
       "import": "./dist/esm/ort.wasm.min.js",
       "require": "./dist/cjs/ort.wasm.min.js",
+      "types": "./types.d.ts",
       "default": "./dist/ort.wasm.min.js"
     },
     "./wasm-core": {
       "import": "./dist/esm/ort.wasm-core.min.js",
       "require": "./dist/cjs/ort.wasm-core.min.js",
+      "types": "./types.d.ts",
       "default": "./dist/ort.wasm-core.min.js"
     },
     "./webgl": {
       "import": "./dist/esm/ort.webgl.min.js",
       "require": "./dist/cjs/ort.webgl.min.js",
+      "types": "./types.d.ts",
       "default": "./dist/ort.webgl.min.js"
     },
     "./webgpu": {
       "import": "./dist/esm/ort.webgpu.min.js",
       "require": "./dist/cjs/ort.webgpu.min.js",
+      "types": "./types.d.ts",
       "default": "./dist/ort.webgpu.min.js"
     },
     "./training": {
       "import": "./dist/esm/ort.training.wasm.min.js",
       "require": "./dist/cjs/ort.training.wasm.min.js",
+      "types": "./types.d.ts",
       "default": "./dist/ort.training.wasm.min.js"
     }
   },

diff --git a/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_gpt.h b/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_gpt.h
@@ -258,7 +258,7 @@ Status BeamSearchGpt<T>::Execute(const FeedsFetchesManager* init_run_feeds_fetch
     cpu_state.sequences.InitDevice(beam_state.sequences_device);
     ORT_RETURN_IF_ERROR(this->device_copy_int32_func_(beam_state.sequences_device.subspan(0, beam_state.sequences_device.size() / 2),
                                                       cpu_state.sequences_space.subspan(0, cpu_state.sequences_space.size() / 2),
-                                                      nullptr,
+                                                      this->ort_stream_,
                                                       DeviceCopyDirection::hostToDevice));
   }
 

diff --git a/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_t5.h b/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_t5.h
@@ -214,7 +214,7 @@ Status BeamSearchT5<T>::Execute(const FeedsFetchesManager& encoder_feeds_fetches
     cpu_state.sequences.InitDevice(beam_state.sequences_device);
     ORT_RETURN_IF_ERROR(this->device_copy_int32_func_(beam_state.sequences_device.subspan(0, beam_state.sequences_device.size() / 2),
                                                       cpu_state.sequences_space.subspan(0, cpu_state.sequences_space.size() / 2),
-                                                      nullptr,
+                                                      this->ort_stream_,
                                                       DeviceCopyDirection::hostToDevice));
   }
 

diff --git a/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_whisper.h b/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_whisper.h
@@ -134,8 +134,8 @@ Status BeamSearchWhisper<T>::Execute(const FeedsFetchesManager& encoder_feeds_fe
     TensorShape no_speech_probs_shape{parameters->batch_size};
     Tensor* no_speech_probs = this->context_.Output(parameters->no_speech_probs_output_id, no_speech_probs_shape);
     if (no_speech_probs && no_speech_probs->MutableData<T>()) {
-      ORT_ENFORCE(parameters->no_speech_token >= 0 && parameters->no_speech_token < parameters->vocab_size,
-                  "no_speech_token id out of range, it is ", parameters->no_speech_token,
+      ORT_ENFORCE(parameters->no_speech_token_id >= 0 && parameters->no_speech_token_id < parameters->vocab_size,
+                  "no_speech_token_id is out of range, it is ", parameters->no_speech_token_id,
                   ", vocab_size is ", parameters->vocab_size);
       this->parameters_->no_speech_probs = (void*)no_speech_probs->MutableData<T>();
     }
@@ -226,7 +226,7 @@ Status BeamSearchWhisper<T>::Execute(const FeedsFetchesManager& encoder_feeds_fe
     cpu_state.sequences.InitDevice(beam_state.sequences_device);
     ORT_RETURN_IF_ERROR(this->device_copy_int32_func_(beam_state.sequences_device.subspan(0, beam_state.sequences_device.size() / 2),
                                                       cpu_state.sequences_space.subspan(0, cpu_state.sequences_space.size() / 2),
-                                                      nullptr,
+                                                      this->ort_stream_,
                                                       DeviceCopyDirection::hostToDevice));
   }
 

diff --git a/onnxruntime/contrib_ops/cpu/transformers/beam_search_parameters.cc b/onnxruntime/contrib_ops/cpu/transformers/beam_search_parameters.cc
@@ -141,7 +141,13 @@ void WhisperBeamSearchParameters::ParseFromAttributes(const OpKernelInfo& info)
   model_type = static_cast<int>(info.GetAttrOrDefault<int64_t>("model_type", IGenerationParameters::kModelTypeWhisper));
   ORT_ENFORCE(model_type == IGenerationParameters::kModelTypeWhisper);
 
-  no_speech_token = static_cast<int>(info.GetAttrOrDefault<int64_t>("no_speech_token", -1LL));
+  // Token ids are defined below in the order that they appear in the tokenizer
+  translate_token_id = static_cast<int>(info.GetAttrOrDefault<int64_t>("translate_token_id", -1LL));
+  transcribe_token_id = static_cast<int>(info.GetAttrOrDefault<int64_t>("transcribe_token_id", -1LL));
+  start_of_lm_token_id = static_cast<int>(info.GetAttrOrDefault<int64_t>("start_of_lm_token_id", -1LL));
+  no_speech_token_id = static_cast<int>(info.GetAttrOrDefault<int64_t>("no_speech_token_id", -1LL));
+  no_timestamps_token_id = static_cast<int>(info.GetAttrOrDefault<int64_t>("no_timestamps_token_id", -1LL));
+  beginning_timestamp_token_id = static_cast<int>(info.GetAttrOrDefault<int64_t>("beginning_timestamp_token_id", -1LL));
   cross_qk_layer_head_input_id = 12;
   extra_decoding_ids_input_id = 13;
   cross_qk_output_id = 3;

diff --git a/onnxruntime/contrib_ops/cpu/transformers/generation_shared.h b/onnxruntime/contrib_ops/cpu/transformers/generation_shared.h
@@ -180,7 +180,14 @@ struct IGenerationParameters {
   // Parameters for whisper model
   bool decoder_output_cross_qk = false;
   gsl::span<const int32_t> extra_decoding_ids;
-  int32_t no_speech_token = -1;
+
+  // Token ids are defined below in the order that they appear in the tokenizer
+  int32_t translate_token_id = -1;
+  int32_t transcribe_token_id = -1;
+  int32_t start_of_lm_token_id = -1;
+  int32_t no_speech_token_id = -1;
+  int32_t no_timestamps_token_id = -1;
+  int32_t beginning_timestamp_token_id = -1;
   void* no_speech_probs = nullptr;
 
   int cross_qk_layer_head_input_id = -1;