add int8 wo quantization

Signed-off-by: Yuening Li <[email protected]>

Author: Yuening-wa

docs/source/guides/_compress_quantized_models.rst

@@ -32,7 +32,7 @@ After PTQ, the model can be compressed with the following code:
 Initialize HF models with compressed weights for lower memory usage
 ===================================================================
 
-When working with large language models, memory constraints can be a significant challenge. ModelOpt provides a workflow for initaializing HF models with compressed weights across multiple GPUs to dramatically reduce memory usage.
+When working with large language models, memory constraints can be a significant challenge. ModelOpt provides a workflow for initializing HF models with compressed weights across multiple GPUs to dramatically reduce memory usage.
 
 For quantized formats like NVFP4, you can reduce memory usage by up to 4x compared to FP16/BF16 models. One limitation is that this workflow only works with max calibration algorithm.
 

examples/llm_ptq/hf_ptq.py

@@ -58,6 +58,7 @@
 QUANT_CFG_CHOICES: dict[str, dict[str, Any]] = {
     "int8": mtq.INT8_DEFAULT_CFG,
     "int8_sq": mtq.INT8_SMOOTHQUANT_CFG,
+    "int8_wo": mtq.INT8_WEIGHT_ONLY_CFG,
     "fp8": mtq.FP8_DEFAULT_CFG,
     "int4_awq": mtq.INT4_AWQ_CFG,
     "w4a8_awq": mtq.W4A8_AWQ_BETA_CFG,
@@ -87,7 +88,16 @@ def auto_quantize(
     if args.export_fmt == "hf":
         assert all(
             qformat
-            in ["fp8", "int4_awq", "nvfp4", "nvfp4_awq", "w4a8_awq", "fp8_pb_wo", "w4a8_mxfp4_fp8"]
+            in [
+                "fp8",
+                "int8_wo",
+                "int4_awq",
+                "nvfp4",
+                "nvfp4_awq",
+                "w4a8_awq",
+                "fp8_pb_wo",
+                "w4a8_mxfp4_fp8",
+            ]
             for qformat in qformat_list
         ), (
             "One or more quantization formats provided are not supported for unified checkpoint export"
@@ -215,6 +225,8 @@ def main(args):
             assert (
                 args.qformat
                 in [
+                    "int8",
+                    "int8_wo",
                     "int4_awq",
                     "fp8",
                     "nvfp4",

examples/llm_ptq/scripts/huggingface_example.sh

@@ -45,9 +45,9 @@ if [ "$EXPORT_FORMAT" = "hf" ]; then
     IFS=","
     for qformat in $QFORMAT; do
         case $qformat in
-        fp16 | bf16 | fp8 | fp8_pc_pt | fp8_pb_wo | int4_awq | nvfp4 | nvfp4_awq | w4a8_awq | w4a8_mxfp4_fp8) ;;
+        fp16 | bf16 | fp8 | fp8_pc_pt | fp8_pb_wo | int8_wo | int4_awq | nvfp4 | nvfp4_awq | w4a8_awq | w4a8_mxfp4_fp8) ;;
         *)
-            echo "Unsupported quant argument: Expected one of: [fp16, bf16, fp8, fp8_pc_pt, fp8_pb_wo, int4_awq, nvfp4, nvfp4_awq, w4a8_awq, w4a8_mxfp4_fp8]" >&2
+            echo "Unsupported quant argument: Expected one of: [fp16, bf16, fp8, fp8_pc_pt, fp8_pb_wo, int8_wo, int4_awq, nvfp4, nvfp4_awq, w4a8_awq, w4a8_mxfp4_fp8]" >&2
             exit 1
             ;;
         esac
@@ -74,9 +74,9 @@ esac
 IFS=","
 for qformat in $QFORMAT; do
     case $qformat in
-    fp8 | fp8_pc_pt | fp8_pb_wo | int8_sq | int4_awq | w4a8_awq | fp16 | bf16 | nvfp4 | nvfp4_awq | w4a8_mxfp4_fp8) ;;
+    fp8 | fp8_pc_pt | fp8_pb_wo | int8_wo | int8_sq | int4_awq | w4a8_awq | fp16 | bf16 | nvfp4 | nvfp4_awq | w4a8_mxfp4_fp8) ;;
     *)
-        echo "Unknown quant argument: Expected one of: [fp8, fp8_pc_pt, fp8_pb_wo, int8_sq, int4_awq, w4a8_awq, fp16, bf16, nvfp4, nvfp4_awq, w4a8_mxfp4_fp8]" >&2
+        echo "Unknown quant argument: Expected one of: [fp8, fp8_pc_pt, fp8_pb_wo, int8_wo, int8_sq, int4_awq, w4a8_awq, fp16, bf16, nvfp4, nvfp4_awq, w4a8_mxfp4_fp8]" >&2
         exit 1
         ;;
     esac

modelopt/torch/export/model_config.py

@@ -29,6 +29,7 @@
 QUANTIZATION_NONE = None
 QUANTIZATION_FP8 = "fp8"
 QUANTIZATION_INT8_SQ = "int8_sq"
+QUANTIZATION_INT8_WO = "int8_wo"
 QUANTIZATION_INT4_AWQ = "int4_awq"
 QUANTIZATION_W4A8_AWQ = "w4a8_awq"
 QUANTIZATION_NVFP4 = "nvfp4"
@@ -200,7 +201,7 @@ def weights_scaling_factor(self):
             return None
 
         if self.q.weights_scaling_factor.numel() != 1:
-            # for NVFP4, Int4 AWQ and Int8 SQ case, we concatenate the
+            # for NVFP4, Int4 AWQ, Int8 SQ and Int8_WO case, we concatenate the
             # q_weight_scaling_factor, k_weight_scaling_factor, v_weight_scaling_factor
             if self.q.weights_scaling_factor.dtype == torch.float8_e4m3fn:
                 # For NVFP4, we recompute the weights_scaling_factor using max weights_scaling_factor2

modelopt/torch/export/quant_utils.py

@@ -50,6 +50,7 @@
     QUANTIZATION_FP8_PC_PT,
     QUANTIZATION_INT4_AWQ,
     QUANTIZATION_INT8_SQ,
+    QUANTIZATION_INT8_WO,
     QUANTIZATION_MXFP4,
     QUANTIZATION_NONE,
     QUANTIZATION_NVFP4,
@@ -419,7 +420,10 @@ def _get_quantization_from_layer(layer, quantizer_attr_names: QuantizerAttrNames
             return QUANTIZATION_INT4_AWQ
 
         if weight_quantizer.num_bits == 8:
-            return QUANTIZATION_INT8_SQ
+            if input_quantizer is not None and input_quantizer.is_enabled:
+                return QUANTIZATION_INT8_SQ
+            else:
+                return QUANTIZATION_INT8_WO
 
         if weight_quantizer.num_bits == (4, 3):
             if weight_quantizer.block_sizes:
@@ -578,6 +582,8 @@ def process_layer_quant_config(layer_config_dict):
                 "has_zero_point": False,
                 "pre_quant_scale": True,
             }
+        elif v == "int8_wo":
+            layer_config = {"quant_algo": "W8A16"}
         elif v == "int8_sq":
             layer_config = {"quant_algo": "W8A8_SQ_PER_CHANNEL"}
         elif v == "nvfp4":
@@ -696,7 +702,7 @@ def to_quantized_weight(
             return (weight / weights_scaling_factor.unsqueeze(-1)).to(torch.float8_e4m3fn)
         return (weight / weights_scaling_factor).to(torch.float8_e4m3fn)
 
-    if quantization == QUANTIZATION_INT8_SQ:
+    if quantization in [QUANTIZATION_INT8_SQ, QUANTIZATION_INT8_WO]:
         return (weight / weights_scaling_factor[:, None]).round().clamp(-128, 127).to(torch.int8)
 
     if quantization == QUANTIZATION_FP8_PB_WO:
@@ -748,7 +754,7 @@ def from_quantized_weight(
             torch_dtype
         )
 
-    if quantization == QUANTIZATION_INT8_SQ:
+    if quantization in [QUANTIZATION_INT8_SQ, QUANTIZATION_INT8_WO]:
         return weight.to(torch_dtype) * weights_scaling_factor[:, None].to(torch_dtype)
 
     raise NotImplementedError(f"quantization format {quantization} not supported")

modelopt/torch/quantization/config.py

@@ -169,6 +169,15 @@
     "algorithm": "max",
 }
 
+INT8_WEIGHT_ONLY_CFG = {
+    "quant_cfg": {
+        "*weight_quantizer": {"num_bits": 8, "axis": 0},
+        "*input_quantizer": {"enable": False},
+        **_default_disabled_quantizer_cfg,
+    },
+    "algorithm": "max",
+}
+
 INT8_SMOOTHQUANT_CFG = {
     "quant_cfg": {
         "*weight_quantizer": {"num_bits": 8, "axis": 0},
@@ -573,6 +582,7 @@
     "INT4_AWQ_CFG",
     "INT4_BLOCKWISE_WEIGHT_ONLY_CFG",
     "INT8_DEFAULT_CFG",
+    "INT8_WO_CFG",
     "INT8_SMOOTHQUANT_CFG",
     "MXFP4_DEFAULT_CFG",
     "MXFP8_DEFAULT_CFG",

tests/gpu/torch/export/test_export.py

@@ -57,6 +57,7 @@
     FP8_DEFAULT_CFG,
     INT4_AWQ_CFG,
     INT8_SMOOTHQUANT_CFG,
+    INT8_WEIGHT_ONLY_CFG,
     NVFP4_AWQ_LITE_CFG,
     NVFP4_DEFAULT_CFG,
     W4A8_AWQ_BETA_CFG,
@@ -323,6 +324,7 @@ def test_adjust_attn_amax_values(
     ("config", "expected_block_size"),
     [
         (FP8_DEFAULT_CFG, 0),
+        (INT8_WEIGHT_ONLY_CFG, 0),
         (INT8_SMOOTHQUANT_CFG, 0),
         (NVFP4_DEFAULT_CFG, 16),
         (NVFP4_AWQ_LITE_CFG, 16),