2

config.json

@@ -34,7 +34,7 @@
   "num_mel_bins": 80,
   "pad_token_id": 50257,
   "scale_embedding": false,
-  "torch_dtype": "float16",
+  "torch_dtype": "float32",
   "transformers_version": "4.26.0.dev0",
   "use_cache": false,
   "vocab_size": 51865

run.sh

@@ -12,7 +12,7 @@ deepspeed run_speech_recognition_seq2seq_streaming-farsipal.py \
 	--torch_compile_mode="reduce-overhead" \
 	--torch_compile_mode="max-autotune" \
 	--max_steps="10000" \
-	--resume_from_checkpoint="checkpoint-4000" \
+	--resume_from_checkpoint="checkpoint-6000" \
 	--output_dir="./" \
 	--per_device_train_batch_size="32" \
 	--gradient_accumulation_steps="1" \

runs/Dec17_22-33-02_129-146-176-120/events.out.tfevents.1671317315.129-146-176-120.1102971.0

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b7e9d6ba2e6b67e4ffa281e2c6f41fce3b8ee132cd4e45f3b483c061ad9f4395
-size 10922
+oid sha256:2060292f5085682e66c0f37431096c4f396d2f9f5dc66f34e4c69ccc255ceb27
+size 17520