Add wint4afp8 & fix fp quant bugs

configs/quantization/backend/sglang/fp8/awq_fp8.yml

@@ -26,14 +26,15 @@ eval:
     inference_per_block: False
 quant:
     method: Awq
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/backend/sglang/fp8/awq_fp8_static.yml

@@ -26,14 +26,15 @@ eval:
     inference_per_block: False
 quant:
     method: Awq
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_tensor
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/backend/sglang/fp8/gptq_fp8.yml

@@ -26,14 +26,15 @@ eval:
     inference_per_block: False
 quant:
     method: GPTQ
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/backend/sglang/fp8/rtn_fp8.yml

@@ -17,13 +17,14 @@ eval:
     inference_per_block: False
 quant:
     method: RTN
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         bit: e4m3
         symmetric: True
         granularity: per_token

configs/quantization/backend/sglang/fp8/smoothquant_fp8.yml

@@ -22,14 +22,15 @@ eval:
     seq_len: 2048
 quant:
     method: SmoothQuant
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/backend/vllm/fp8/awq_fp8.yml

@@ -26,14 +26,15 @@ eval:
     inference_per_block: False
 quant:
     method: Awq
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/backend/vllm/fp8/awq_fp8_static.yml

@@ -26,14 +26,15 @@ eval:
     inference_per_block: False
 quant:
     method: Awq
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_tensor
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/backend/vllm/fp8/gptq_fp8.yml

@@ -26,14 +26,15 @@ eval:
     inference_per_block: False
 quant:
     method: GPTQ
-    quant_type: float_quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/backend/vllm/fp8/rtn_fp8.yml

@@ -17,13 +17,14 @@ eval:
     inference_per_block: False
 quant:
     method: RTN
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         bit: e4m3
         symmetric: True
         granularity: per_token

configs/quantization/backend/vllm/fp8/smoothquant_fp8.yml

@@ -22,14 +22,15 @@ eval:
     seq_len: 2048
 quant:
     method: SmoothQuant
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True
         granularity: per_channel
         use_qtorch: True
     act:
+        quant_type: float-quant
         # Support ["e4m3", "e5m2"]
         bit: e4m3
         symmetric: True

configs/quantization/methods/FP_Quant/awq_we2m1a16_g128.yml

@@ -25,14 +25,15 @@ eval:
     inference_per_block: False
 quant:
     method: Awq
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e2m1
         symmetric: False
         granularity: per_group
         group_size: 128
         use_qtorch: True
     special:
+        quant_type: float-quant
         trans: True
         # The options for "trans_version" include "v1" and "v2".
         trans_version: v2

configs/quantization/methods/FP_Quant/gptq_we2m1a16_g128.yml

@@ -26,14 +26,15 @@ eval:
     inference_per_block: False
 quant:
     method: GPTQ
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e2m1
         symmetric: True
         granularity: per_group
         group_size: 128
         use_qtorch: True
     special:
+        quant_type: float-quant
         actorder: True
         static_groups: False
         percdamp: 0.01

configs/quantization/methods/FP_Quant/rtn_we2m1a16_g128.yml

@@ -16,8 +16,8 @@ eval:
     inference_per_block: False
 quant:
     method: RTN
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e2m1
         symmetric: True
         granularity: per_group

configs/quantization/methods/FP_Quant/rtn_we2m1ae2m1.yml

@@ -16,12 +16,13 @@ eval:
     inference_per_block: False
 quant:
     method: RTN
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e2m1
         symmetric: True
         granularity: per_channel
     act:
+        quant_type: float-quant
         bit: e2m1
         symmetric: True
         granularity: per_token

configs/quantization/methods/FP_Quant/rtn_we4m3ae4m3.yml

@@ -1,27 +1,28 @@
 base:
     seed: &seed 42
 model:
-    type: model_type
-    path: model path
+    type: Llama
+    path: /mnt/nvme1/yongyang/models/llama2-7b
     torch_dtype: auto
-eval:
-    eval_pos: [pretrain, fake_quant]
-    name: wikitext2
-    download: False
-    path: eval data path
-    seq_len: 2048
-    # For 7B / 13B model eval, bs can be set to "1", and inference_per_block can be set to "False".
-    # For 70B model eval, bs can be set to "20", and inference_per_block can be set to "True".
-    bs: 1
-    inference_per_block: False
+# eval:
+#     eval_pos: [pretrain, fake_quant]
+#     name: wikitext2
+#     download: False
+#     path: /mnt/nvme0/yongyang/llm_datasets/llmc/eval/wikitext2
+#     seq_len: 2048
+#     # For 7B / 13B model eval, bs can be set to "1", and inference_per_block can be set to "False".
+#     # For 70B model eval, bs can be set to "20", and inference_per_block can be set to "True".
+#     bs: 1
+#     inference_per_block: False
 quant:
     method: RTN
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e4m3
         symmetric: True
         granularity: per_channel
     act:
+        quant_type: float-quant
         bit: e4m3
         symmetric: True
         granularity: per_token

configs/quantization/methods/FP_Quant/rtn_we5m2ae5m2.yml

@@ -16,12 +16,13 @@ eval:
     inference_per_block: False
 quant:
     method: RTN
-    quant_type: float-quant
     weight:
+        quant_type: float-quant
         bit: e5m2
         symmetric: True
         granularity: per_channel
     act:
+        quant_type: float-quant
         bit: e5m2
         symmetric: True
         granularity: per_token

configs/quantization/methods/RTN/rtn_w_a_wint4afp8.yml

@@ -0,0 +1,40 @@
+base:
+    seed: &seed 42
+model:
+    type: Llama
+    path: /mnt/nvme1/yongyang/models/llama2-7b
+    torch_dtype: auto
+eval:
+    eval_pos: [pretrain, fake_quant]
+    name: wikitext2
+    download: False
+    path: /mnt/nvme0/yongyang/llm_datasets/llmc/eval/wikitext2
+    seq_len: 2048
+    # For 7B / 13B model eval, bs can be set to "1", and inference_per_block can be set to "False".
+    # For 70B model eval, bs can be set to "20", and inference_per_block can be set to "True".
+    bs: 1
+    inference_per_block: False
+quant:
+    method: RTN
+    weight:
+        bit: 48
+        bit4:
+            symmetric: False
+            granularity: per_group
+            group_size: 128
+            scales_bit: 8
+            scales_symmetric: True
+            zeros_bit: 8
+            zeros_symmetric: True
+        bit8:
+            symmetric: True
+            granularity: per_channel
+            int_range: [-120, 120]
+    act:
+        quant_type: float-quant
+        bit: e4m3
+        symmetric: True
+        granularity: per_token
+save:
+    save_fake: False
+    save_path: /path/to/save/

configs/quantization/methods/RTN/rtn_w_a_wint4aint8.yml

@@ -17,7 +17,6 @@ eval:
 quant:
     method: RTN
     weight:
-        quant_type: int-quant
         bit: 48
         bit4:
             symmetric: False
@@ -32,7 +31,6 @@ quant:
             granularity: per_channel
             int_range: [-120, 120]
     act:
-        quant_type: int-quant
         bit: 8
         symmetric: True
         granularity: per_token

llmc/compression/quantization/quant.py

@@ -37,7 +37,7 @@ def __init__(self, bit, symmetric, granularity, **kwargs):
         # hist config
         self.bins = self.kwargs.get('bins', 2048)
         self.hist_threshold = self.kwargs.get('hist_threshold', 1)
-        self.dst_nbins = 2**bit
+        self.dst_nbins = 2**bit if isinstance(bit, int) else None
         self.upsample_rate = (
             16  # used to reduce quantization errors when upscaling histogram
         )