[Examples] Use more robust splits (#1544)

kylesayrs · web-flow · commit 4f40c6d2a8dc · 2025-06-13T17:51:14.000Z
## Purpose ##
* Make tests more robust to custom model ids that don't have a "/" in
the name
* Many users replace the model ids with local paths. These local paths
do not necessarily have "/"s in them, and will fail at the very end of
compression, which is decidedly a bad time
* Also account for cases where users pass paths to directories whose
names end with "/"

## Changes ##
* `model_id.split("/")[1]` -&gt; `model_id.rstrip("/").split("/")[-1]`

## Testing ##
* `grep -r -i 'MODEL_ID.split("/")\[1\]' src examples/ tests`

---------

Signed-off-by: Kyle Sayers &lt;kylesayrs@gmail.com&gt;
diff --git a/examples/awq/llama_example.py b/examples/awq/llama_example.py
@@ -72,6 +72,6 @@ def tokenize(sample):
 print("==========================================\n\n")
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[-1] + "-awq-asym"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-awq-asym"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/awq/qwen3_moe_example.py b/examples/awq/qwen3_moe_example.py
@@ -77,6 +77,6 @@ def tokenize(sample):
 print("==========================================\n\n")
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[-1] + "-awq-sym"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-awq-sym"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/big_models_with_accelerate/cpu_offloading_fp8.py b/examples/big_models_with_accelerate/cpu_offloading_fp8.py
@@ -4,7 +4,7 @@
 from llmcompressor.modifiers.quantization import QuantizationModifier
 
 MODEL_ID = "meta-llama/Meta-Llama-3-70B-Instruct"
-OUTPUT_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+OUTPUT_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 
 # Load model
 # Note: device_map="auto" will offload to CPU if not enough space on GPU.
diff --git a/examples/big_models_with_accelerate/mult_gpus_int8_device_map.py b/examples/big_models_with_accelerate/mult_gpus_int8_device_map.py
@@ -68,7 +68,7 @@ def tokenize(sample):
     ),
 ]
 
-SAVE_DIR = MODEL_ID.split("/")[1] + "-INT8"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-INT8"
 
 oneshot(
     model=model,
diff --git a/examples/big_models_with_accelerate/multi_gpu_int8.py b/examples/big_models_with_accelerate/multi_gpu_int8.py
@@ -5,7 +5,7 @@
 from llmcompressor.modifiers.quantization import GPTQModifier
 
 MODEL_ID = "meta-llama/Meta-Llama-3-70B-Instruct"
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W8A8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W8A8-Dynamic"
 
 # 1) Load model (device_map="auto" with shard the model over multiple GPUs!).
 model = AutoModelForCausalLM.from_pretrained(
diff --git a/examples/multimodal_audio/whisper_example.py b/examples/multimodal_audio/whisper_example.py
@@ -106,6 +106,6 @@ def data_collator(batch):
 # and it was a great thing for what it was at the time but it's not a passive house
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/gemma3_example.py b/examples/multimodal_vision/gemma3_example.py
@@ -74,6 +74,6 @@ def data_collator(batch):
 print("==========================================")
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/idefics3_example.py b/examples/multimodal_vision/idefics3_example.py
@@ -115,6 +115,6 @@ def tokenize(sample):
 print("==========================================")
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/llava_example.py b/examples/multimodal_vision/llava_example.py
@@ -70,6 +70,6 @@ def data_collator(batch):
 print("==========================================")
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/mistral3_example.py b/examples/multimodal_vision/mistral3_example.py
@@ -84,6 +84,6 @@ def data_collator(batch):
 print("==========================================")
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/mllama_example.py b/examples/multimodal_vision/mllama_example.py
@@ -70,6 +70,6 @@ def data_collator(batch):
 print("==========================================")
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/phi3_vision_example.py b/examples/multimodal_vision/phi3_vision_example.py
@@ -98,6 +98,6 @@ def data_collator(batch):
 print("==========================================")
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/pixtral_example.py b/examples/multimodal_vision/pixtral_example.py
@@ -76,6 +76,6 @@ def data_collator(batch):
 print("==========================================")
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/qwen2_vl_example.py b/examples/multimodal_vision/qwen2_vl_example.py
@@ -130,6 +130,6 @@ def data_collator(batch):
 
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/multimodal_vision/qwen_2_5_vl_example.py b/examples/multimodal_vision/qwen_2_5_vl_example.py
@@ -124,6 +124,6 @@ def data_collator(batch):
 
 
 # Save to disk compressed.
-SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_kv_cache/README.md b/examples/quantization_kv_cache/README.md
@@ -127,7 +127,7 @@ print(tokenizer.decode(output[0]))
 Save the quantized model:
 
 ```python
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-KV"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
 ```
diff --git a/examples/quantization_kv_cache/gemma2_fp8_kv_example.py b/examples/quantization_kv_cache/gemma2_fp8_kv_example.py
@@ -98,6 +98,6 @@ def process_and_tokenize(example):
 print("==========================================\n\n")
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-KV"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_kv_cache/llama3_fp8_kv_example.py b/examples/quantization_kv_cache/llama3_fp8_kv_example.py
@@ -96,6 +96,6 @@ def process_and_tokenize(example):
 print("==========================================\n\n")
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-KV"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_kv_cache/phi3.5_fp8_kv_example.py b/examples/quantization_kv_cache/phi3.5_fp8_kv_example.py
@@ -96,6 +96,6 @@ def process_and_tokenize(example):
 print("==========================================\n\n")
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-KV"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w4a16/README.md b/examples/quantization_w4a16/README.md
@@ -101,7 +101,7 @@ oneshot(
 )
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
 ```
diff --git a/examples/quantization_w4a16/llama3_example.py b/examples/quantization_w4a16/llama3_example.py
@@ -75,6 +75,6 @@ def tokenize(sample):
 print("==========================================\n\n")
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W4A16-G128"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W4A16-G128"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w4a16_fp4/llama3_example.py b/examples/quantization_w4a16_fp4/llama3_example.py
@@ -28,6 +28,6 @@
 
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-NVFP4A16"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4A16"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w4a4_fp4/llama3_example.py b/examples/quantization_w4a4_fp4/llama3_example.py
@@ -76,6 +76,6 @@ def tokenize(sample):
 
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-NVFP4"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_fp8/README.md b/examples/quantization_w8a8_fp8/README.md
@@ -65,7 +65,7 @@ recipe = QuantizationModifier(
 oneshot(model=model, recipe=recipe)
 
 # Save the model.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 model.save_pretrained(SAVE_DIR)
 tokenizer.save_pretrained(SAVE_DIR)
 ```
diff --git a/examples/quantization_w8a8_fp8/gemma2_example.py b/examples/quantization_w8a8_fp8/gemma2_example.py
@@ -37,6 +37,6 @@
 print("==========================================")
 
 # 4) Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_fp8/llama3.2_vision_example.py b/examples/quantization_w8a8_fp8/llama3.2_vision_example.py
@@ -32,6 +32,6 @@
 print("==========================================")
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_fp8/llama3_example.py b/examples/quantization_w8a8_fp8/llama3_example.py
@@ -30,6 +30,6 @@
 print("==========================================")
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 model.save_pretrained(SAVE_DIR)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_fp8/llava1.5_example.py b/examples/quantization_w8a8_fp8/llava1.5_example.py
@@ -32,6 +32,6 @@
 print("==========================================")
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_fp8/qwen2vl_example.py b/examples/quantization_w8a8_fp8/qwen2vl_example.py
@@ -32,6 +32,6 @@
 print("==========================================")
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_fp8/whisper_example.py b/examples/quantization_w8a8_fp8/whisper_example.py
@@ -41,6 +41,6 @@
 print("==========================================")
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 processor.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_int8/README.md b/examples/quantization_w8a8_int8/README.md
@@ -106,7 +106,7 @@ oneshot(
 )
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W8A8-Dynamic-Per-Token"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W8A8-Dynamic-Per-Token"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
 ```
diff --git a/examples/quantization_w8a8_int8/gemma2_example.py b/examples/quantization_w8a8_int8/gemma2_example.py
@@ -77,6 +77,6 @@ def tokenize(sample):
 print("==========================================")
 
 # 5) Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-INT8"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-INT8"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantization_w8a8_int8/llama3_example.py b/examples/quantization_w8a8_int8/llama3_example.py
@@ -80,6 +80,6 @@ def tokenize(sample):
 print("==========================================\n\n")
 
 # Save to disk compressed.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W8A8-Dynamic-Per-Token"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W8A8-Dynamic-Per-Token"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantizing_moe/deepseek_moe_w4a16.py b/examples/quantizing_moe/deepseek_moe_w4a16.py
@@ -95,7 +95,7 @@ def tokenize(sample):
     )
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W4A16"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W4A16"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
 
diff --git a/examples/quantizing_moe/deepseek_moe_w8a8_fp8.py b/examples/quantizing_moe/deepseek_moe_w8a8_fp8.py
@@ -92,6 +92,6 @@ def tokenize(sample):
     )
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantizing_moe/deepseek_moe_w8a8_int8.py b/examples/quantizing_moe/deepseek_moe_w8a8_int8.py
@@ -105,6 +105,6 @@ def tokenize(sample):
     )
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-W8A8"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W8A8"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantizing_moe/mixtral_moe_w8a8_fp8.py b/examples/quantizing_moe/mixtral_moe_w8a8_fp8.py
@@ -60,6 +60,6 @@
     )
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = f"{MODEL_ID.split('/')[-1]}-FP8"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/quantizing_moe/qwen_moe_w4a16.py b/examples/quantizing_moe/qwen_moe_w4a16.py
@@ -88,6 +88,6 @@ def tokenize(sample):
 print("==========================================")
 
 # Save to disk in compressed-tensors format.
-SAVE_DIR = MODEL_ID.split("/")[1] + "-quantized.w4a16"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-quantized.w4a16"
 model.save_pretrained(SAVE_DIR, save_compressed=True)
 tokenizer.save_pretrained(SAVE_DIR)
diff --git a/examples/sparse_2of4_quantization_fp8/llama3_8b_2of4.py b/examples/sparse_2of4_quantization_fp8/llama3_8b_2of4.py
@@ -48,7 +48,7 @@ def get_recipe(fp8_enabled):
             targets=[r"re:model.layers.\d*$"],
         )
     ]
-    save_dir = MODEL_ID.split("/")[1] + "2of4-sparse"
+    save_dir = MODEL_ID.rstrip("/").split("/")[-1] + "2of4-sparse"
 
     if fp8_enabled:
         base_recipe.append(
@@ -58,7 +58,9 @@ def get_recipe(fp8_enabled):
                 scheme="FP8_DYNAMIC",
             )
         )
-        save_dir = MODEL_ID.split("/")[1] + "2of4-W8A8-FP8-Dynamic-Per-Token"
+        save_dir = (
+            MODEL_ID.rstrip("/").split("/")[-1] + "2of4-W8A8-FP8-Dynamic-Per-Token"
+        )
 
         # check that asymmetric quantization is not being used
         q_scheme = base_recipe[1].scheme

Original file line number	Diff line number	Diff line change
`@@ -68,7 +68,7 @@ def tokenize(sample):`
`68`	`68`	`),`
`69`	`69`	`]`
`70`	`70`
`71`		`-SAVE_DIR = MODEL_ID.split("/")[1] + "-INT8"`
	`71`	`+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-INT8"`
`72`	`72`
`73`	`73`	`oneshot(`
`74`	`74`	`model=model,`