whisper.rn 0.4.0-rc.10 → 0.4.0-rc.12

package/cpp/whisper.h

@@ -523,8 +523,8 @@ extern "C" {
         bool detect_language;
 
         // common decoding parameters:
-        bool suppress_blank;    // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L89
-        bool suppress_non_speech_tokens; // ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
+        bool suppress_blank; // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L89
+        bool suppress_nst;   // non-speech tokens, ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
 
         float temperature;      // initial decoding temperature, ref: https://ai.stackexchange.com/a/32478
         float max_initial_ts;   // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L97
@@ -535,7 +535,7 @@ extern "C" {
         float temperature_inc;
         float entropy_thold;    // similar to OpenAI's "compression_ratio_threshold"
         float logprob_thold;
-        float no_speech_thold;  // TODO: not implemented
+        float no_speech_thold;
 
         struct {
             int best_of;    // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L264
@@ -666,6 +666,9 @@ extern "C" {
 
     WHISPER_API void whisper_log_set(wsp_ggml_log_callback log_callback, void * user_data);
 
+    // Get the no_speech probability for the specified segment
+    WHISPER_API float whisper_full_get_segment_no_speech_prob           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API float whisper_full_get_segment_no_speech_prob_from_state(struct whisper_state * state, int i_segment);
 #ifdef __cplusplus
 }
 #endif

package/lib/commonjs/version.json

@@ -1 +1 @@
-{"version":"1.7.2"}
+{"version":"1.7.4"}

package/lib/module/version.json

@@ -1 +1 @@
-{"version":"1.7.2"}
+{"version":"1.7.4"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.rn",
-  "version": "0.4.0-rc.10",
+  "version": "0.4.0-rc.12",
   "description": "React Native binding of whisper.cpp",
   "main": "lib/commonjs/index",
   "module": "lib/module/index",
@@ -15,6 +15,7 @@
     "ios",
     "cpp/*.*",
     "cpp/coreml/*.*",
+    "cpp/amx/*.*",
     "*.podspec",
     "!lib/typescript/example",
     "!ios/build",

package/src/version.json

@@ -1 +1 @@
-{"version":"1.7.2"}
+{"version":"1.7.4"}

package/whisper-rn.podspec

@@ -2,7 +2,7 @@ require "json"
 
 package = JSON.parse(File.read(File.join(__dir__, "package.json")))
 base_ld_flags = "-framework Accelerate -framework Foundation -framework Metal -framework MetalKit"
-base_compiler_flags = "-DWSP_GGML_USE_ACCELERATE -Wno-shorten-64-to-32"
+base_compiler_flags = "-DWSP_GGML_USE_CPU -DWSP_GGML_USE_ACCELERATE -Wno-shorten-64-to-32"
 folly_compiler_flags = "-DFOLLY_NO_CONFIG -DFOLLY_MOBILE=1 -DFOLLY_USE_LIBCPP=1 -Wno-comma"
 
 # Use base_optimizer_flags = "" for debug builds
@@ -42,7 +42,7 @@ Pod::Spec.new do |s|
   s.pod_target_xcconfig = {
     "OTHER_LDFLAGS" => base_ld_flags,
     "OTHER_CFLAGS" => base_optimizer_flags,
-    "OTHER_CPLUSPLUSFLAGS" => base_optimizer_flags
+    "OTHER_CPLUSPLUSFLAGS" => base_optimizer_flags + " -std=c++17"
   }
 
   # Don't install the dependencies when we run `pod install` in the old architecture.

package/cpp/README.md

@@ -1,4 +0,0 @@
-# Note
-
-- Only `rn-*` are the specific files for this project, others are sync from [whisper.cpp](https://github.com/ggerganov/whisper.cpp).
-- We can update the native source by using the [bootstrap](../scripts/bootstrap.sh) script.

package/cpp/ggml-aarch64.c

@@ -1,129 +0,0 @@
-#define WSP_GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml-aarch64.h"
-#include "ggml-impl.h"
-#include "ggml-quants.h"
-#include <assert.h>
-
-#define UNUSED WSP_GGML_UNUSED
-
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
-    block_q4_0x4 out;
-
-    for (int i = 0; i < 4; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    const int end = QK4_0 * 2 / blck_size_interleave;
-
-    if (blck_size_interleave == 8) {
-        const uint64_t xor_mask = 0x8888888888888888ULL;
-        for (int i = 0; i < end; ++i) {
-            int src_id = i % 4;
-            int src_offset = (i / 4) * blck_size_interleave;
-            int dst_offset = i * blck_size_interleave;
-
-            uint64_t elems;
-            // Using memcpy to avoid unaligned memory accesses
-            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-            elems ^= xor_mask;
-            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
-        }
-    } else if (blck_size_interleave == 4) {
-        const uint32_t xor_mask = 0x88888888;
-        for (int i = 0; i < end; ++i) {
-            int src_id = i % 4;
-            int src_offset = (i / 4) * blck_size_interleave;
-            int dst_offset = i * blck_size_interleave;
-
-            uint32_t elems;
-            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
-            elems ^= xor_mask;
-            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
-        }
-    } else {
-        WSP_GGML_ASSERT(false);
-    }
-
-    return out;
-}
-
-// interleave 8 block_q4_0s in blocks of blck_size_interleave
-// returns an interleaved block_q4_0x8
-// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
-// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
-    block_q4_0x8 out;
-
-    for (int i = 0; i < 8; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    const int end = QK4_0 * 4 / blck_size_interleave;
-    const uint64_t xor_mask = 0x8888888888888888ULL;
-
-    for (int i = 0; i < end; ++i) {
-        int src_id = i % 8;
-        int src_offset = (i / 8) * blck_size_interleave;
-        int dst_offset = i * blck_size_interleave;
-
-        uint64_t elems;
-        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        elems ^= xor_mask;
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
-    }
-
-    return out;
-}
-
-static size_t wsp_quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
-    assert(n_per_row % QK4_0 == 0);
-    const int nb = n_per_row / QK4_0;
-
-    void * out_ptr = NULL;
-    if (nrows_interleaved == 8) {
-        out_ptr = (block_q4_0x8 *) dst;
-    }
-    else if (nrows_interleaved == 4) {
-        out_ptr = (block_q4_0x4 *) dst;
-    }
-    assert(nrows_interleaved <= 8);
-    block_q4_0 dst_tmp[8];
-
-    for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
-
-        for (int64_t x = 0; x < nb; x++) {
-
-            for (int i  = 0; i < nrows_interleaved; i++ ) {
-                wsp_quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
-            }
-
-            if (nrows_interleaved == 8) {
-                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave);
-                out_ptr = (block_q4_0x8 *) out_ptr + 1;
-            }
-            else if (nrows_interleaved == 4) {
-                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave);
-                out_ptr = (block_q4_0x4 *) out_ptr + 1;
-            }
-        }
-    }
-
-    return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
-}
-
-size_t wsp_quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return wsp_quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
-}
-
-size_t wsp_quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return wsp_quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
-}
-
-size_t wsp_quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return wsp_quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
-}

package/cpp/ggml-aarch64.h

@@ -1,19 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t wsp_quantize_q4_0_4x4(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t wsp_quantize_q4_0_4x8(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t wsp_quantize_q4_0_8x8(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-#ifdef __cplusplus
-}
-#endif
-

package/cpp/ggml-backend.cpp.rej

@@ -1,12 +0,0 @@
-@@ -578,8 +578,11 @@
-         register_backend(wsp_ggml_backend_cuda_reg());
- #endif
- #ifdef WSP_GGML_USE_METAL
-+#include <TargetConditionals.h>
-+#if !TARGET_OS_SIMULATOR
-         register_backend(wsp_ggml_backend_metal_reg());
- #endif
-+#endif
- #ifdef WSP_GGML_USE_SYCL
-         register_backend(wsp_ggml_backend_sycl_reg());
- #endif