fix

flexflow · Feb 18, 2025 · 6256118 · 6256118
1 parent 6d38983
commit 6256118
Show file tree

Hide file tree

Showing 3 changed files with 14 additions and 14 deletions.
diff --git a/tests/fine_grained_alignment_test.sh b/tests/fine_grained_alignment_test.sh
@@ -10,6 +10,9 @@ PP_DEGREE=${PP_DEGREE:-1}
 CACHE_PATH=${FF_CACHE_PATH:-"~/.cache/flexflow"}
 NUM_STEPS=${NUM_STEPS:-2}
 FULL_PRECISION=${FULL_PRECISION:-true}
+FUSION=${FUSION:-true}
+export LEGION_BACKTRACE=1
+export FF_DEBG_NO_WEIGHTS=1
 
 cleanup() {
     eval rm -rf "${CACHE_PATH}/debug" ./fine_grained_alignment_config.json ./inference/output/fine_grained_alignment_test_ff.txt ./inference/output/fine_grained_alignment_test_hf.txt
@@ -28,11 +31,6 @@ echo '["Three tips for staying healthy are: "]' > ./inference/prompt/test.json
 # Create output folder
 mkdir -p ./inference/output
 
-# Enable backtrace in case we run into a segfault or assertion failure
-export LEGION_BACKTRACE=1
-export FF_DEBG_NO_WEIGHTS=1
-FUSION=true
-
 
 # Check if the Python code executed successfully
 if ! PROMPT_LENGTH=$(python -c "
@@ -67,12 +65,12 @@ json_config=$(cat <<-END
         "data_parallelism_degree": 1,
         "tensor_parallelism_degree": ${TP_DEGREE},
         "pipeline_parallelism_degree": ${PP_DEGREE},
-        "inference_debugging": ${FULL_PRECISION},
+        "inference_debugging": true,
         "fusion": ${FUSION},
         "refresh_cache": false,
         "llm_model": "${MODEL_NAME}",
         "cache_path": "${CACHE_PATH}",
-        "full_precision": false,
+        "full_precision": ${FULL_PRECISION},
         "prompt": "./inference/prompt/test.json",
         "max_length": $MAX_LENGTH,
         "output_file": "./inference/output/fine_grained_alignment_test_ff.txt"

diff --git a/tests/inference/inference_alignment_test.py b/tests/inference/inference_alignment_test.py
@@ -289,9 +289,9 @@ def compare(hf_tensor, ff_tensor, label="", additional_ff_tensor=None, tolerance
                 ff_qproj_out = np.concatenate((ff_qproj_out, ff_qproj_out_), axis=0)
                 ff_kproj_out = np.concatenate((ff_kproj_out, ff_kproj_out_), axis=0)
                 ff_vproj_out = np.concatenate((ff_vproj_out, ff_vproj_out_), axis=0)
-            compare_loaded_tensors(hf_q_proj_out.T, ff_qproj_out)
-            compare_loaded_tensors(hf_k_proj_out.T, ff_kproj_out)
-            compare_loaded_tensors(hf_v_proj_out.T, ff_vproj_out)
+            compare_loaded_tensors(hf_q_proj_out.T, ff_qproj_out, label=f"Q proj {i} output")
+            compare_loaded_tensors(hf_k_proj_out.T, ff_kproj_out, label=f"K proj {i} output")
+            compare_loaded_tensors(hf_v_proj_out.T, ff_vproj_out, label=f"V proj {i} output")
             ff_tensor_name = f"layers.{i}.layers.{i}.self_attn"
             input_comparison = TensorComparisonIdxs(hf_tensor_type="input", ff_tensor_type="input", hf_tensor_idx=0, ff_tensor_idx=0)
             ff_attn_tensor_in = get_ff_tensor(

diff --git a/tests/peft/alignment/align_test_utils.py b/tests/peft/alignment/align_test_utils.py
@@ -415,7 +415,7 @@ def load_hf_tensor(filename: str):
     return hf_tensor
 
 
-def compare_loaded_tensors(hf_tensor, ff_tensor, tolerance=1e-2):
+def compare_loaded_tensors(hf_tensor, ff_tensor, tolerance=1e-2, label=""):
     """Check whether a Huggingface and a FlexFlow tensors, both loaded to memory in the form of a numpy array, are equal
 
     Args:
@@ -425,13 +425,15 @@ def compare_loaded_tensors(hf_tensor, ff_tensor, tolerance=1e-2):
     """
     assert hf_tensor.shape == ff_tensor.shape
     mismatches = []
+    len_hf_tensor = hf_tensor.flatten().shape[0]
     if not np.allclose(hf_tensor, ff_tensor, atol=tolerance):
-        print(f"mismatch between hf_tensor and ff_tensor")
+        mismatches = np.where(~np.isclose(hf_tensor.squeeze(), ff_tensor.squeeze(), atol=tolerance))[0]
+        label = label + ": " if label else ""
+        pct_mismatch = len(mismatches) / len_hf_tensor
+        print(f"{label} {pct_mismatch*100:.3}% mismatch between hf_tensor and ff_tensor")
         print(f"HF: {hf_tensor.squeeze()}\nFF: {ff_tensor.squeeze()}")
         print(np.isclose(hf_tensor.squeeze(), ff_tensor.squeeze(), atol=tolerance))
-        mismatches = np.where(~np.isclose(hf_tensor.squeeze(), ff_tensor.squeeze(), atol=tolerance))[0]
         # print(mismatches)
-    len_hf_tensor = hf_tensor.flatten().shape[0]
     assert len(mismatches) <= 0.05 * len_hf_tensor
     print("Ok!")