Dev/1.18 (#8)

* Update to 1.1.8.0. * Add MACs. * Power should be measured in the host not inside the container. --------- Co-authored-by: joonhyung.lee <joonhyung.lee@navercorp.com>
NAVER-INTEL-Co-Lab · Oct 30, 2024 · a825a28 · a825a28
1 parent 8645f55
commit a825a28
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 6 deletions.
diff --git a/docker-compose.yaml b/docker-compose.yaml
@@ -106,6 +106,6 @@ services:
       dockerfile: Dockerfile
       args:
         OS: ${OS:-ubuntu22.04}
-        PYTORCH_VERSION: ${PYTORCH_VERSION:-2.3.1}
-        SYNAPSE_VERSION: ${SYNAPSE_VERSION:-1.17.0}
+        PYTORCH_VERSION: ${PYTORCH_VERSION:-2.4.0}
+        SYNAPSE_VERSION: ${SYNAPSE_VERSION:-1.18.0}
         IMAGE_TAG: ${IMAGE_TAG:-latest}
diff --git a/environment.yaml b/environment.yaml
@@ -16,4 +16,4 @@ dependencies: # Use conda packages if possible.
   - pip  # For `pip` dependencies that are not available in conda.
   - pip:
       # Modify the version tag as necessary.
-      - git+https://github.com/huggingface/optimum-habana@v1.13.2
+      - git+https://github.com/huggingface/optimum-habana@v1.14.1
diff --git a/prefill/llama.py b/prefill/llama.py
@@ -10,7 +10,8 @@
     --num_steps 32
 ```
     
-For power measurements, use one of the following commands.
+For power measurements, use one of the following commands on the host.
+The container will likely not have `ipmitool` available.
 
 ```bash
 sudo ipmitool dcmi power reading 5_sec
@@ -83,7 +84,7 @@ def measure(
 ) -> dict:
     config = AutoConfig.from_pretrained(model_name, torch_dtype=torch.bfloat16)
 
-    flops = approx_llama_forward_macs(
+    macs = approx_llama_forward_macs(
         num_decoder_blocks=config.num_hidden_layers,
         sequence_length=seq_len,
         vocabulary_size=config.vocab_size,
@@ -94,7 +95,7 @@ def measure(
         gated_ffn_act=True,
     )
 
-    flops *= 2 * batch_size  # 1 MAC is approximately 2 FLOPs.
+    flops = macs * 2 * batch_size  # 1 MAC is approximately 2 FLOPs.
     device = torch.device("hpu")  # HPUs do not have numbers, unlike NVIDIA GPUs.
     x = torch.zeros(size=(batch_size, seq_len), dtype=torch.int64, device=device)
 
@@ -136,6 +137,7 @@ def measure(
         "Model Min TFLOPS": min(tfps),
         "Model Max TFLOPS": max(tfps),
         "Model STDEV TFLOPS": stdev(tfps),
+        "Forward MAC Count": macs,
     }
     return info