whisper : make large version explicit + fix data size units (#1493)

diff --git a/Makefile b/Makefile
index 23dbb4f3a7f6866d03b6a8316271577fdcea97cf..5358f0633eeadfed0f83229366f4096065b8bfdd 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -418,9 +418,9 @@ samples:
 .PHONY: medium
 .PHONY: large-v1
 .PHONY: large-v2
-.PHONY: large
+.PHONY: large-v3
 
-tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3: main
        bash ./models/download-ggml-model.sh $@
        @echo ""
        @echo "==============================================="

diff --git a/README.md b/README.md
index e2c247c47f35925162631dde49ed17322fcf6a1e..ff97ab103526d17c6e4501a35bf3a5d93ed83074 100644 (file)
--- a/README.md
+++ b/README.md
@@ -231,18 +231,18 @@ make medium.en
 make medium
 make large-v1
 make large-v2
-make large
+make large-v3
 ```
 
 ## Memory usage
 
-| Model  | Disk   | Mem     | SHA                                        |
-| ---    | ---    | ---     | ---                                        |
-| tiny   |  75 MB | ~125 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
-| base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
-| small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
-| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
-| large  | 2.9 GB | ~3.3 GB | `ad82bf6a9043ceed055076d0fd39f5f186ff8062` |
+| Model  | Disk    | Mem      |
+| ---    | ---     | ---      |
+| tiny   |  75 MiB | ~273 MB |
+| base   | 142 MiB | ~388 MB |
+| small  | 466 MiB | ~852 MB |
+| medium | 1.5 GiB | ~2.1 GB |
+| large  | 2.9 GiB | ~3.9 GB |
 
 ## Quantization
 

diff --git a/bindings/go/examples/go-model-download/main.go b/bindings/go/examples/go-model-download/main.go
index d3e45c28ea3ecc5ef438f7d6fd76695a000b6e67..3522d881717b42246a5c41c2da0b3af6575b9262 100644 (file)
--- a/bindings/go/examples/go-model-download/main.go
+++ b/bindings/go/examples/go-model-download/main.go
@@ -24,7 +24,7 @@ const (
 
 var (
        // The models which will be downloaded, if no model is specified as an argument
-       modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large-v2", "ggml-large"}
+       modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large-v2", "ggml-large-v3"}
 )
 
 var (

diff --git a/examples/livestream.sh b/examples/livestream.sh
index d86a7c601d5ce39dd882d575701cfa16c5cc7d04..fbeb6dbccfdb92584df07f07e6fef9050c396934 100755 (executable)
--- a/examples/livestream.sh
+++ b/examples/livestream.sh
@@ -48,7 +48,7 @@ if [ -n "$3" ]; then
 fi
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" )
 
 # list available models
 function list_models {

diff --git a/examples/twitch.sh b/examples/twitch.sh
index 77b618dde9b9e1566069163f382baa4c79771db8..0403fea90153b4ba65f64d5402984bdabcbe1d3e 100755 (executable)
--- a/examples/twitch.sh
+++ b/examples/twitch.sh
@@ -21,7 +21,7 @@ help()
     echo "Usage: ./twitch.sh -s [step] -m [model] -t [threads] [url]"
     echo "options:"
     echo "-s       Step in seconds (default is $step)."
-    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large-v2' 'large' (default is '$model')."
+    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large-v2' 'large-v3' (default is '$model')."
     echo "-t       Number of threads to use."
     echo "-h       Print this help page."
     echo

diff --git a/extra/convert-all.sh b/extra/convert-all.sh
index c9638079c9296300b6549d16aa7ee04942b7e8b4..ff765c92d9e7e11ab5db18eee30121ad93429d61 100755 (executable)
--- a/extra/convert-all.sh
+++ b/extra/convert-all.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
 
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" )
 
 for model in "${models[@]}"; do
     python3 models/convert-pt-to-ggml.py ~/.cache/whisper/$model.pt ../whisper models/

diff --git a/ggml-metal.m b/ggml-metal.m
index 3d22b0b27e444f2a1995a7e729ee7171faa189a3..4fe9cc487d29f6eb2cac0c4fb0a172d6b6d28136 100644 (file)
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -346,9 +346,9 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     }
 
     GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
-    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6);
     if (ctx->device.maxTransferRate != 0) {
-        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6);
     } else {
         GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
     }
@@ -541,11 +541,11 @@ bool ggml_metal_add_buffer(
             ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
             if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1e6);
                 return false;
             }
 
-            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1e6);
 
             ++ctx->n_buffers;
         } else {
@@ -565,11 +565,11 @@ bool ggml_metal_add_buffer(
                 ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
                 if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1e6);
                     return false;
                 }
 
-                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1e6, i);
                 if (i + size_step < size) {
                     GGML_METAL_LOG_INFO("\n");
                 }
@@ -580,8 +580,8 @@ bool ggml_metal_add_buffer(
 
 #if TARGET_OS_OSX
         GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
-                ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
-                ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+                ctx->device.currentAllocatedSize / 1e6,
+                ctx->device.recommendedMaxWorkingSetSize / 1e6);
 
         if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
             GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
@@ -589,7 +589,7 @@ bool ggml_metal_add_buffer(
             GGML_METAL_LOG_INFO("\n");
         }
 #else
-        GGML_METAL_LOG_INFO(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1e6);
 #endif
     }
 

diff --git a/models/README.md b/models/README.md
index b12f2d22d63fbc422b43d6a0ca2fb15a8421584d..de4a13fe2825f45497f7c2e03bf48046bed5f4b0 100644 (file)
--- a/models/README.md
+++ b/models/README.md
@@ -39,19 +39,19 @@ https://huggingface.co/ggerganov/whisper.cpp/tree/main
 
 ## Available models
 
-| Model     | Disk   | Mem     | SHA                                        |
-| ---       | ---    | ---     | ---                                        |
-| tiny      |  75 MB | ~390 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
-| tiny.en   |  75 MB | ~390 MB | `c78c86eb1a8faa21b369bcd33207cc90d64ae9df` |
-| base      | 142 MB | ~500 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
-| base.en   | 142 MB | ~500 MB | `137c40403d78fd54d454da0f9bd998f78703390c` |
-| small     | 466 MB | ~1.0 GB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
-| small.en  | 466 MB | ~1.0 GB | `db8a495a91d927739e50b3fc1cc4c6b8f6c2d022` |
-| medium    | 1.5 GB | ~2.6 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
-| medium.en | 1.5 GB | ~2.6 GB | `8c30f0e44ce9560643ebd10bbe50cd20eafd3723` |
-| large-v1  | 2.9 GB | ~4.7 GB | `b1caaf735c4cc1429223d5a74f0f4d0b9b59a299` |
-| large-v2  | 2.9 GB | ~4.7 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
-| large     | 2.9 GB | ~4.7 GB | `ad82bf6a9043ceed055076d0fd39f5f186ff8062` |
+| Model     | Disk    | SHA                                        |
+| ---       | ---     | ---                                        |
+| tiny      |  75 MiB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
+| tiny.en   |  75 MiB | `c78c86eb1a8faa21b369bcd33207cc90d64ae9df` |
+| base      | 142 MiB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
+| base.en   | 142 MiB | `137c40403d78fd54d454da0f9bd998f78703390c` |
+| small     | 466 MiB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
+| small.en  | 466 MiB | `db8a495a91d927739e50b3fc1cc4c6b8f6c2d022` |
+| medium    | 1.5 GiB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
+| medium.en | 1.5 GiB | `8c30f0e44ce9560643ebd10bbe50cd20eafd3723` |
+| large-v1  | 2.9 GiB | `b1caaf735c4cc1429223d5a74f0f4d0b9b59a299` |
+| large-v2  | 2.9 GiB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| large-v3  | 2.9 GiB | `ad82bf6a9043ceed055076d0fd39f5f186ff8062` |
 
 ## Model files for testing purposes
 

diff --git a/models/convert-h5-to-coreml.py b/models/convert-h5-to-coreml.py
index 3887c22a7e40c8dcddc14a7ee2fcc34f3d81772e..57341ab0995f7e8274cfa552c151e318a085fbd7 100644 (file)
--- a/models/convert-h5-to-coreml.py
+++ b/models/convert-h5-to-coreml.py
@@ -78,14 +78,14 @@ def convert_hf_whisper(hf_model_name_or_path: str, whisper_state_path: str):
 # Ported from models/convert-whisper-to-coreml.py
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model-name", type=str, help="name of model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large, large-v1, large-v2)", required=True)
+    parser.add_argument("--model-name", type=str, help="name of model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large-v1, large-v2, large-v3)", required=True)
     parser.add_argument("--model-path", type=str, help="path to the model (e.g. if published on HuggingFace: Oblivion208/whisper-tiny-cantonese)", required=True)
     parser.add_argument("--encoder-only", type=bool, help="only convert encoder", default=False)
     parser.add_argument("--quantize",     type=bool, help="quantize weights to F16", default=False)
     parser.add_argument("--optimize-ane", type=bool, help="optimize for ANE execution (currently broken)", default=False)
     args = parser.parse_args()
 
-    if args.model_name not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large", "large-v1", "large-v2"]:
+    if args.model_name not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large-v1", "large-v2", "large-v3"]:
         raise ValueError("Invalid model name")
 
     pt_target_path = f"models/hf-{args.model_name}.pt"

diff --git a/models/convert-whisper-to-coreml.py b/models/convert-whisper-to-coreml.py
index 7e09f5ba6a900eb0079bcf3f87ee7c95f61b4e15..fd7191abcb27e7f2108f0767d08611a254452804 100644 (file)
--- a/models/convert-whisper-to-coreml.py
+++ b/models/convert-whisper-to-coreml.py
@@ -296,13 +296,13 @@ def convert_decoder(hparams, model, quantize=False):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model", type=str, help="model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large, large-v1, large-v2)", required=True)
+    parser.add_argument("--model", type=str, help="model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large-v1, large-v2, large-v3)", required=True)
     parser.add_argument("--encoder-only", type=bool, help="only convert encoder", default=False)
     parser.add_argument("--quantize",     type=bool, help="quantize weights to F16", default=False)
     parser.add_argument("--optimize-ane", type=bool, help="optimize for ANE execution (currently broken)", default=False)
     args = parser.parse_args()
 
-    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "small.en-tdrz", "medium", "medium.en", "large", "large-v1", "large-v2"]:
+    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "small.en-tdrz", "medium", "medium.en", "large-v1", "large-v2", "large-v3"]:
         raise ValueError("Invalid model name")
 
     whisper = load_model(args.model).cpu()

diff --git a/models/convert-whisper-to-openvino.py b/models/convert-whisper-to-openvino.py
index 88e03ff7550541b200a647735c5e0af960820f54..1a4ad304df015fb5bf199e6827aea7dee241e4b7 100644 (file)
--- a/models/convert-whisper-to-openvino.py
+++ b/models/convert-whisper-to-openvino.py
@@ -38,10 +38,10 @@ def convert_encoder(hparams, encoder, mname):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model", type=str, help="model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large, large-v1, large-v2)", required=True)
+    parser.add_argument("--model", type=str, help="model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large-v1, large-v2, large-v3)", required=True)
     args = parser.parse_args()
 
-    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large", "large-v1", "large-v2"]:
+    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large-v1", "large-v2", "large-v3"]:
         raise ValueError("Invalid model name")
 
     whisper = load_model(args.model).cpu()

diff --git a/models/download-coreml-model.sh b/models/download-coreml-model.sh
index 95739dbf9957bdf9c64b86d30f448ecbb10a72cf..9e67a15002f67bf40fd3614b094e7723107a4974 100755 (executable)
--- a/models/download-coreml-model.sh
+++ b/models/download-coreml-model.sh
@@ -19,7 +19,7 @@ function get_script_path() {
 models_path="$(get_script_path)"
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" )
 
 # list available models
 function list_models {

diff --git a/models/download-ggml-model.cmd b/models/download-ggml-model.cmd
index fc279967dac418a636f044b509a69e94877a32ee..4d21531d2e1d32346a17f9b2ffd053ebeee1119e 100644 (file)
--- a/models/download-ggml-model.cmd
+++ b/models/download-ggml-model.cmd
@@ -8,7 +8,7 @@ popd
 set argc=0
 for %%x in (%*) do set /A argc+=1
 
-set models=tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large
+set models=tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3
 
 if %argc% neq 1 (
   echo.

diff --git a/models/download-ggml-model.sh b/models/download-ggml-model.sh
index ea68da8936df288b629342e01f211b30166756db..9c0119c3df9b5f66b649f9c9ecb00f1636fd8bb2 100755 (executable)
--- a/models/download-ggml-model.sh
+++ b/models/download-ggml-model.sh
@@ -22,7 +22,7 @@ function get_script_path() {
 models_path="$(get_script_path)"
 
 # Whisper models
-models=( 
+models=(
     "tiny.en"
     "tiny"
     "tiny-q5_1"
@@ -42,7 +42,7 @@ models=(
     "medium.en-q5_0"
     "large-v1"
     "large-v2"
-    "large"
+    "large-v3"
     "large-q5_0"
 )
 

diff --git a/tests/run-tests.sh b/tests/run-tests.sh
index bf062dd6b8d29ba3e66dc76fcd13815975af8baa..a9606a1f8a826e12e30f7839ce967726de99ee28 100755 (executable)
--- a/tests/run-tests.sh
+++ b/tests/run-tests.sh
@@ -19,7 +19,7 @@
 cd `dirname $0`
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" )
 
 # list available models
 function list_models {

diff --git a/whisper.cpp b/whisper.cpp
index a3e0fbd007769130e235cb33f661aaf5fe51558a..acedee0d08cd5ef0180adfb07339a5fc2dadcfec 100644 (file)
--- a/whisper.cpp
+++ b/whisper.cpp
@@ -1522,7 +1522,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
         model.buffer = ggml_backend_alloc_buffer(wctx.backend, size_main);
 
-        WHISPER_LOG_INFO("%s: %8s buffer size = %8.2f MB\n", __func__, ggml_backend_name(wctx.backend), size_main / 1024.0 / 1024.0);
+        WHISPER_LOG_INFO("%s: %8s buffer size = %8.2f MB\n", __func__, ggml_backend_name(wctx.backend), size_main / 1e6);
     }
 
     ggml_allocr * alloc = ggml_allocr_new_from_buffer(model.buffer);
@@ -1637,12 +1637,12 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                 ggml_backend_tensor_set(tensor, read_buf.data(), 0, ggml_nbytes(tensor));
             }
 
-            //printf("%48s - [%5d, %5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype), ggml_nbytes(tensor)/1024.0/1024.0);
+            //printf("%48s - [%5d, %5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype), ggml_nbytes(tensor)/1e6);
             total_size += ggml_nbytes(tensor);
             model.n_loaded++;
         }
 
-        WHISPER_LOG_INFO("%s: model size    = %7.2f MB\n", __func__, total_size/1024.0/1024.0);
+        WHISPER_LOG_INFO("%s: model size    = %7.2f MB\n", __func__, total_size/1e6);
 
         if (model.n_loaded == 0) {
             WHISPER_LOG_WARN("%s: WARN no tensors loaded from model file - assuming empty model for testing\n", __func__);
@@ -2027,11 +2027,11 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
     ////////////////////////////////////////////////////////////////////////////
 
     //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
-    //        ggml_used_mem(ctx0)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
+    //        ggml_used_mem(ctx0)/1e6,
+    //        wstate.get_buf_max_mem(0)/1e6,
+    //        wstate.get_buf_max_mem(1)/1e6,
+    //        wstate.get_buf_max_mem(2)/1e6,
+    //        wstate.get_buf_max_mem(3)/1e6);
 
     ggml_free(ctx0);
 
@@ -2613,11 +2613,11 @@ static bool whisper_decode_internal(
 
     if (batch.n_tokens > 1) {
         //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
-        //        ggml_used_mem(ctx0)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
+        //        ggml_used_mem(ctx0)/1e6,
+        //        wstate.get_buf_max_mem(0)/1e6,
+        //        wstate.get_buf_max_mem(1)/1e6,
+        //        wstate.get_buf_max_mem(2)/1e6,
+        //        wstate.get_buf_max_mem(3)/1e6);
     }
 
     if (batch.n_tokens == 1) {
@@ -3057,7 +3057,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
 
     {
         const size_t memory_size = ggml_nbytes(state->kv_self.k) + ggml_nbytes(state->kv_self.v);
-        WHISPER_LOG_INFO("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
+        WHISPER_LOG_INFO("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1e6);
     }
 
     if (!kv_cache_init(ctx->model.hparams, state->kv_cross, ctx->backend, ctx->itype, ctx->model.hparams.n_audio_ctx)) {
@@ -3068,7 +3068,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
 
     {
         const size_t memory_size = ggml_nbytes(state->kv_cross.k) + ggml_nbytes(state->kv_cross.v);
-        WHISPER_LOG_INFO("%s: kv cross size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
+        WHISPER_LOG_INFO("%s: kv cross size = %7.2f MB\n", __func__, memory_size / 1e6);
     }
 
 #ifdef WHISPER_USE_COREML
@@ -3110,7 +3110,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
                     return whisper_build_graph_conv(*ctx, *state, 0);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (conv)   = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_conv) / 1024.0 / 1024.0);
+        WHISPER_LOG_INFO("%s: compute buffer (conv)   = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_conv) / 1e6);
     }
 
     // encoder allocator
@@ -3120,7 +3120,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
                     return whisper_build_graph_encoder(*ctx, *state);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (encode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_encode) / 1024.0 / 1024.0);
+        WHISPER_LOG_INFO("%s: compute buffer (encode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_encode) / 1e6);
     }
 
     // cross allocator
@@ -3130,7 +3130,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
                     return whisper_build_graph_cross(*ctx, *state);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (cross)  = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_cross) / 1024.0 / 1024.0);
+        WHISPER_LOG_INFO("%s: compute buffer (cross)  = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_cross) / 1e6);
     }
 
     // decoder allocator
@@ -3148,7 +3148,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
                     return whisper_build_graph_decoder(*ctx, *state, state->batch);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (decode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_decode) / 1024.0 / 1024.0);
+        WHISPER_LOG_INFO("%s: compute buffer (decode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_decode) / 1e6);
     }
 
     whisper_allocr_graph_realloc(state->alloc_conv,   ctx->backend);
@@ -6072,8 +6072,8 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
     size_t n    = 20;
     size_t arr  = n_threads > 0 ? 1024llu : n_threads; // trick to avoid compiler optimizations
 
-    // 1GB MB array
-    const size_t size = arr*1024llu*1024llu;
+    // 1GB array
+    const size_t size = arr*1e9;
 
     // single-thread
     {
@@ -6099,7 +6099,7 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
             src[rand() % size] = rand() % 256;
         }
 
-        snprintf(strbuf, sizeof(strbuf), "memcpy: %.2f GB/s (1 thread)\n", (double) (n*size)/(tsum*1024llu*1024llu*1024llu));
+        snprintf(strbuf, sizeof(strbuf), "memcpy: %.2f GB/s (1 thread)\n", (double) (n*size)/(tsum*1e9));
         s += strbuf;
 
         // needed to prevent the compiler from optimizing the memcpy away