rename and add tests

vllm-project · Aug 30, 2024 · bfa3ff6 · bfa3ff6
1 parent 143ec4b
commit bfa3ff6
Show file tree

Hide file tree

Showing 4 changed files with 26 additions and 2 deletions.
diff --git a/...rs/compression/configs/actorder_1.1b.yaml → ...pression/configs/actorder_group_1.1b.yaml b/...rs/compression/configs/actorder_1.1b.yaml → ...pression/configs/actorder_group_1.1b.yaml
@@ -1,5 +1,5 @@
 cadence: "nightly"
 test_type: "regression"
 model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
-new_recipe: "tests/llmcompressor/transformers/compression/recipes/new_quant_actorder.yaml"
+new_recipe: "tests/llmcompressor/transformers/compression/recipes/new_quant_actorder_group.yaml"
 ppl_threshold: 20
diff --git a/tests/llmcompressor/transformers/compression/configs/actorder_weight_1.1b.yaml b/tests/llmcompressor/transformers/compression/configs/actorder_weight_1.1b.yaml
@@ -0,0 +1,5 @@
+cadence: "nightly"
+test_type: "regression"
+model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+new_recipe: "tests/llmcompressor/transformers/compression/recipes/new_quant_actorder_weight.yaml"
+ppl_threshold: 20
diff --git a/...mpression/recipes/new_quant_actorder.yaml → ...ion/recipes/new_quant_actorder_group.yaml b/...mpression/recipes/new_quant_actorder.yaml → ...ion/recipes/new_quant_actorder_group.yaml
@@ -10,7 +10,7 @@ test_stage:
                         symmetric: False
                         strategy: "group"
                         group_size: 128
-                        actorder: True
+                        actorder: "group"
                     input_activations: null
                     output_activations: null
                     targets: ["Linear"]

diff --git a/tests/llmcompressor/transformers/compression/recipes/new_quant_actorder_weight.yaml b/tests/llmcompressor/transformers/compression/recipes/new_quant_actorder_weight.yaml
@@ -0,0 +1,19 @@
+test_stage:
+    quant_modifiers:
+        QuantizationModifier:
+            ignore: ["lm_head", "model.layers.0.mlp.down_proj"]
+            config_groups:
+                group_0:
+                    weights:
+                        num_bits: 4
+                        type: "int"
+                        symmetric: False
+                        strategy: "group"
+                        group_size: 128
+                        actorder: "weight"
+                    input_activations: null
+                    output_activations: null
+                    targets: ["Linear"]
+        GPTQModifier:
+            block_size: 128
+            sequential_update: False