Support FP8 quantize

Author: zhiyuan1i

rwkv_pip_package/src/rwkv/model.py

@@ -233,14 +233,20 @@ def __init__(self, model, strategy):
 
             ss = strategy.split(' ')
             DEVICE = ss[0]
+            self.ffn_fp8 = False
             if ss[1] == 'fp16':
                 DTYPE = torch.half
             elif ss[1] == 'fp32':
                 DTYPE = torch.float32
             elif ss[1] == 'bf16':
                 DTYPE = torch.bfloat16
+            elif ss[1] == 'fp8':
+                DTYPE = torch.half
+                print("FP8 only support with modern CUDA devices, and it only quantize FFN weights")
+                print("The other weights are still in FP16\n\n")
+                self.ffn_fp8 = True
             else:
-                assert False, "currently rwkv7 strategy must be: cuda/cpu fp16/fp32/bf16"
+                assert False, "currently rwkv7 strategy must be: cuda/cpu fp16/fp32/bf16/fp8"
 
             temp_z = torch.load(args.MODEL_NAME + '.pth', map_location='cpu', mmap=True)
 
@@ -257,14 +263,20 @@ def __init__(self, model, strategy):
                 args.n_layer = max(args.n_layer, layer_id+1)
                 tensor = temp_z[k]
                 if 'key.weight' in k or 'value.weight' in k or 'receptance.weight' in k or 'output.weight' in k or 'head.weight' in k:
-                    tensor = tensor.t().contiguous()
+                    tensor = tensor.t()
                 tensor = tensor.squeeze()
                 if k.endswith('att.r_k'): 
                     tensor = tensor.flatten()
-                self.z[k] = tensor.to(DEVICE).to(DTYPE)
+                
+                if ('ffn.key.weight' in k or 'ffn.value.weight' in k) and self.ffn_fp8:
+                    # convert to fp8
+                    tensor, reciprocal = to_float8(tensor)
+                    self.z[k] = tensor.to(DEVICE)
+                    self.z[k+'_scale'] = reciprocal.to(DEVICE)
+                else:
+                    self.z[k] = tensor.contiguous().to(DEVICE).to(DTYPE)
                 del temp_z[k]
-                if keys.index(k) % 5 == 0:
-                    torch.cuda.empty_cache()
+
 
             self.n_embd = args.n_embd
             self.n_layer = args.n_layer
@@ -315,7 +327,10 @@ def forward_one(self, idx:int, state:List[torch.Tensor]):
 
                     xx = F.layer_norm(x, (self.n_embd,), weight=z[bbb+'ln2.weight'], bias=z[bbb+'ln2.bias'])
 
-                    xx, state[i*3+2] = RWKV_x070_CMix_one(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'value.weight'])
+                    if not self.ffn_fp8:
+                        xx, state[i*3+2] = RWKV_x070_CMix_one(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'value.weight'])
+                    else:
+                        xx, state[i*3+2] = RWKV_x070_CMix_one_fp8(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'key.weight_scale'], z[ffn+'value.weight'], z[ffn+'value.weight_scale'])
                     x = x + xx
 
                     # if math.isnan(torch.min(x).item()): print(idx, i)
@@ -347,8 +362,10 @@ def forward_seq(self, idx:List[int], state:List[torch.Tensor], full_output:bool=
                     x = x + xx
 
                     xx = F.layer_norm(x, (self.n_embd,), weight=z[bbb+'ln2.weight'], bias=z[bbb+'ln2.bias'])
-
-                    xx, state[i*3+2] = RWKV_x070_CMix_seq(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'value.weight'])
+                    if not self.ffn_fp8:
+                        xx, state[i*3+2] = RWKV_x070_CMix_seq(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'value.weight'])
+                    else:
+                        xx, state[i*3+2] = RWKV_x070_CMix_seq_fp8(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'key.weight_scale'], z[ffn+'value.weight'], z[ffn+'value.weight_scale'])
                     x = x + xx
 
                 if not full_output: x = x[-1,:]
@@ -442,21 +459,63 @@ def RWKV_x070_TMix_seq(layer_id: int, H:int, N:int, x, x_prev, v_first, state, x
             return (xx * g) @ O_, x[-1,:], state, v_first
 
     ########################################################################################################
+    @MyStatic
+    def to_float8(x):
+        dtype_min = torch.tensor(-448).to(x.device)
+        dtype_max = torch.tensor(448).to(x.device)
+        scale = dtype_max / x.abs().max().clamp(min=1e-12)
+        x_scl_sat = (x * scale).clamp(min=dtype_min, max=dtype_max)
+        return x_scl_sat.to(torch.float8_e4m3fn), scale.float().reciprocal()
+    
+    @MyStatic
+    def to_float8_e5m2(x):
+        dtype_min = torch.tensor(-57344).to(x.device)
+        dtype_max = torch.tensor(57344).to(x.device)
+        abs_max = x.abs().max().clamp(min=1e-12)
+        scale = dtype_max / abs_max
+        x_scl_sat = (x * scale).clamp(min=dtype_min, max=dtype_max)
+        return x_scl_sat.to(torch.float8_e5m2), scale.float().reciprocal() 
 
     @MyStatic
     def RWKV_x070_CMix_one(x, x_prev, x_k, K_, V_):
         xx = x_prev - x
-        k = x + xx * x_k
+        k = torch.addcmul(x, xx, x_k)
         k = torch.relu(k @ K_) ** 2
         return k @ V_, x
 
+    @MyStatic
+    def RWKV_x070_CMix_one_fp8(x, x_prev, x_k, K_, K_scale, V_, V_scale):
+        xx = x_prev - x
+        k = torch.addcmul(x, xx, x_k)
+        k_fp8, k_scale_1 = to_float8(k)
+        k_fp8 = k_fp8.unsqueeze(0)
+        k_new= torch._scaled_mm(k_fp8, K_, k_scale_1, K_scale, out_dtype=x.dtype)
+        k = torch.relu(k_new) ** 2
+        k_fp8, k_scale_2 = to_float8(k)
+        output1 = torch._scaled_mm(k_fp8, V_, k_scale_2, V_scale, out_dtype=x.dtype)
+        return output1.squeeze(0), x
+    
+    
     @MyStatic
     def RWKV_x070_CMix_seq(x, x_prev, x_k, K_, V_):
         xx = torch.cat((x_prev.unsqueeze(0), x[:-1,:])) - x
-        k = x + xx * x_k
+        k = torch.addcmul(x, xx, x_k)
         k = torch.relu(k @ K_) ** 2
         return k @ V_, x[-1,:]
 
+    @MyStatic
+    def RWKV_x070_CMix_seq_fp8(x, x_prev, x_k, K_, K_scale, V_, V_scale):
+        xx = torch.cat((x_prev.unsqueeze(0), x[:-1,:])) - x
+        k = torch.addcmul(x, xx, x_k)
+        k_fp8, k_scale_1 = to_float8(k)
+        k_new = torch._scaled_mm(k_fp8, K_, k_scale_1, K_scale, out_dtype=x.dtype)
+        k = torch.relu(k_new) ** 2
+        k_fp8, k_scale_2 = to_float8(k)
+        output1 = torch._scaled_mm(k_fp8, V_, k_scale_2, V_scale, out_dtype=x.dtype)
+        return output1, x[-1,:]
+
+  
+
 ########################################################################################################
 
 class RWKV(MyModule):