Initial plumbing for skip_zeros.

bitsandbytes/functional.py

@@ -337,7 +337,7 @@ def optimizer_update_32bit(optimizer_name:str, g: Tensor, p: Tensor, state1: Ten
                 beta1: float, eps: float, step: int, lr: float,
                 state2: Tensor=None, beta2: float=0.0,
                 weight_decay: float=0.0, gnorm_scale: float=1.0,
-                unorm_vec: Tensor=None, max_unorm: float=0.0) -> None:
+                unorm_vec: Tensor=None, max_unorm: float=0.0, skip_zeros=False) -> None:
     '''
     Performs an inplace optimizer update with one or two optimizer states.
 
@@ -369,6 +369,12 @@ def optimizer_update_32bit(optimizer_name:str, g: Tensor, p: Tensor, state1: Ten
         Optimizer beta2.
     gnorm_scale : float
         The factor to rescale the gradient to the max clip value.
+    unorm_vec : torch.Tensor
+        The tensor for the update norm.
+    max_unorm : float
+        The maximum update norm relative to the weight norm.
+    skip_zeros : bool
+        Whether to skip zero-valued gradients or not (default: False).
     '''
 
     param_norm = 0.0
@@ -381,11 +387,11 @@ def optimizer_update_32bit(optimizer_name:str, g: Tensor, p: Tensor, state1: Ten
     if g.dtype == torch.float32 and state1.dtype == torch.float32:
         str2optimizer32bit[optimizer_name][0](get_ptr(g), get_ptr(p), get_ptr(state1), get_ptr(state2), get_ptr(unorm_vec), ct.c_float(max_unorm),
                     ct.c_float(param_norm), ct.c_float(beta1), ct.c_float(beta2), ct.c_float(eps), ct.c_float(weight_decay),
-                    ct.c_int32(step), ct.c_float(lr), ct.c_float(gnorm_scale), ct.c_int32(g.numel()))
+                    ct.c_int32(step), ct.c_float(lr), ct.c_float(gnorm_scale), ct.c_bool(skip_zeros), ct.c_int32(g.numel()))
     elif g.dtype == torch.float16 and state1.dtype == torch.float32:
         str2optimizer32bit[optimizer_name][1](get_ptr(g), get_ptr(p), get_ptr(state1), get_ptr(state2), get_ptr(unorm_vec), ct.c_float(max_unorm),
                     ct.c_float(param_norm), ct.c_float(beta1), ct.c_float(beta2), ct.c_float(eps), ct.c_float(weight_decay),
-                    ct.c_int32(step), ct.c_float(lr), ct.c_float(gnorm_scale), ct.c_int32(g.numel()))
+                    ct.c_int32(step), ct.c_float(lr), ct.c_float(gnorm_scale), ct.c_bool(skip_zeros), ct.c_int32(g.numel()))
     else:
         raise ValueError(f'Gradient+optimizer bit data type combination not supported: grad {g.dtype}, optimizer {state1.dtype}')
 
@@ -439,6 +445,10 @@ def optimizer_update_8bit(optimizer_name: str, g: Tensor, p: Tensor, state1: Ten
         Max value for the next Adam update of the second state.
     gnorm_scale : float
         The factor to rescale the gradient to the max clip value.
+    unorm_vec : torch.Tensor
+        The tensor for the update norm.
+    max_unorm : float
+        The maximum update norm relative to the weight norm.
     '''
 
     param_norm = 0.0
@@ -468,19 +478,22 @@ def optimizer_update_8bit(optimizer_name: str, g: Tensor, p: Tensor, state1: Ten
 def optimizer_update_8bit_blockwise(optimizer_name: str, g: Tensor, p: Tensor, state1: Tensor, state2: Tensor,
                 beta1: float, beta2: float, eps: float,
                 step: int, lr: float, qmap1: Tensor, qmap2: Tensor,
-                absmax1: Tensor, absmax2: Tensor, weight_decay: float=0.0, gnorm_scale: float=1.0) -> None:
+                absmax1: Tensor, absmax2: Tensor, weight_decay: float=0.0, gnorm_scale: float=1.0,
+                skip_zeros=False) -> None:
 
 
     if g.dtype == torch.float32 and state1.dtype == torch.uint8:
         str2optimizer8bit_blockwise[optimizer_name][0](get_ptr(p), get_ptr(g), get_ptr(state1), get_ptr(state2),
                     ct.c_float(beta1), ct.c_float(beta2), ct.c_float(eps),
                     ct.c_int32(step), ct.c_float(lr), get_ptr(qmap1), get_ptr(qmap2),
-                    get_ptr(absmax1), get_ptr(absmax2), ct.c_float(weight_decay), ct.c_float(gnorm_scale), ct.c_int32(g.numel()))
+                    get_ptr(absmax1), get_ptr(absmax2), ct.c_float(weight_decay), ct.c_float(gnorm_scale), 
+                    ct.c_bool(skip_zeros), ct.c_int32(g.numel()))
     elif g.dtype == torch.float16 and state1.dtype == torch.uint8:
         str2optimizer8bit_blockwise[optimizer_name][1](get_ptr(p), get_ptr(g), get_ptr(state1), get_ptr(state2),
                     ct.c_float(beta1), ct.c_float(beta2), ct.c_float(eps),
                     ct.c_int32(step), ct.c_float(lr), get_ptr(qmap1), get_ptr(qmap2),
-                    get_ptr(absmax1), get_ptr(absmax2), ct.c_float(weight_decay), ct.c_float(gnorm_scale), ct.c_int32(g.numel()))
+                    get_ptr(absmax1), get_ptr(absmax2), ct.c_float(weight_decay), ct.c_float(gnorm_scale), 
+                    ct.c_bool(skip_zeros), ct.c_int32(g.numel()))
     else:
         raise ValueError(f'Gradient+optimizer bit data type combination not supported: grad {g.dtype}, optimizer {state1.dtype}')
 

bitsandbytes/optim/optimizer.py

@@ -220,6 +220,7 @@ class Optimizer8bit(torch.optim.Optimizer):
         config['percentile_clipping'] = self.args.percentile_clipping
         config['block_wise'] = self.args.block_wise
         config['max_unorm'] = self.args.max_unorm
+        config['skip_zeros'] = self.args.skip_zeros
 
         if (gindex, pindex) in self.mng.index2config:
             config.update(self.mng.index2config[(gindex, pindex)])
@@ -234,7 +235,8 @@ class Optimizer8bit(torch.optim.Optimizer):
 class Optimizer2State(Optimizer8bit):
     def __init__(self, optimizer_name, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8,
             weight_decay=0.0, optim_bits=32, args=None,
-            min_8bit_size=4096, percentile_clipping=100, block_wise=True, max_unorm=0.0):
+            min_8bit_size=4096, percentile_clipping=100, block_wise=True, max_unorm=0.0,
+            skip_zeros=False):
         if not 0.0 <= lr:
             raise ValueError("Invalid learning rate: {}".format(lr))
         if not 0.0 <= eps:
@@ -259,6 +261,7 @@ class Optimizer2State(Optimizer8bit):
             args['percentile_clipping'] = percentile_clipping
             args['block_wise'] = block_wise
             args['max_unorm'] = max_unorm
+            args['skip_zeros'] = skip_zeros
 
             self.args = MockArgs(args)
         else:
@@ -355,7 +358,8 @@ class Optimizer2State(Optimizer8bit):
 class Optimizer1State(Optimizer8bit):
     def __init__(self, optimizer_name, params, lr=1e-3, betas=(0.9, 0.0), eps=1e-8,
             weight_decay=0.0, optim_bits=32, args=None,
-            min_8bit_size=4096, percentile_clipping=100, block_wise=True, max_unorm=0.0):
+            min_8bit_size=4096, percentile_clipping=100, block_wise=True, max_unorm=0.0,
+            skip_zeros=False):
         if not 0.0 <= lr:
             raise ValueError("Invalid learning rate: {}".format(lr))
         if not 0.0 <= eps:
@@ -377,6 +381,7 @@ class Optimizer1State(Optimizer8bit):
             args['percentile_clipping'] = percentile_clipping
             args['block_wise'] = block_wise
             args['max_unorm'] = max_unorm
+            args['skip_zeros'] = skip_zeros
 
             self.args = MockArgs(args)
         else:
@@ -444,7 +449,8 @@ class Optimizer1State(Optimizer8bit):
         if state['state1'].dtype == torch.float:
             F.optimizer_update_32bit(self.optimizer_name, grad, p, state['state1'], config['betas'][0], config['eps'], step, config['lr'],
                     None, 0.0, config['weight_decay'], gnorm_scale,
-                    state['unorm_vec'] if config['max_unorm'] > 0.0 else None, max_unorm=config['max_unorm'])
+                    state['unorm_vec'] if config['max_unorm'] > 0.0 else None, max_unorm=config['max_unorm'],
+                    skip_zeros=False)
 
         elif state['state1'].dtype == torch.uint8 and not config['block_wise']:
             F.optimizer_update_8bit(self.optimizer_name, grad, p, state['state1'], None, config['betas'][0], config['betas'][1],
@@ -457,4 +463,4 @@ class Optimizer1State(Optimizer8bit):
             F.optimizer_update_8bit_blockwise(self.optimizer_name, grad, p, state['state1'], None, config['betas'][0], config['betas'][1],
                           config['eps'],  step, config['lr'],
                           state['qmap1'], None, state['absmax1'], None,
-                          config['weight_decay'], gnorm_scale=gnorm_scale)
+                          config['weight_decay'], gnorm_scale=gnorm_scale, skip_zeros=False)

csrc/kernels.cu

@@ -654,7 +654,7 @@ __launch_bounds__(TH, 1)
 __global__ void kOptimizer32bit2State(T* g, T* p, 
                 float* state1, float* state2, float *unorm, const float max_unorm, const float param_norm,
                 const float beta1, const float beta2, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, const int n)
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
   const int n_full = ((TH*NUM_PER_THREAD)*(n/(TH*NUM_PER_THREAD))) + (n % (TH*NUM_PER_THREAD) == 0 ? 0 : (TH*NUM_PER_THREAD));
@@ -809,7 +809,7 @@ __launch_bounds__(TH, 1)
 __global__ void kOptimizer32bit1State(T *g, T *p, 
                 float *state1, float *unorm, const float max_unorm, const float param_norm,
                 const float beta1, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, const int n)
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
   const int n_full = ((TH*NUM_PER_THREAD)*(n/(TH*NUM_PER_THREAD))) + (n % (TH*NUM_PER_THREAD) == 0 ? 0 : (TH*NUM_PER_THREAD));
@@ -1383,7 +1383,7 @@ kOptimizerStatic8bit2StateBlockwise(T* p, T* __restrict__ const g, unsigned char
                 float* __restrict__ const quantiles1, float* __restrict__ const quantiles2,
                 float* absmax1, float* absmax2, 
                 float weight_decay,
-                const float gnorm_scale, const int n)
+                const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
     //const int n_full = n + (n%BLOCK_SIZE);
@@ -1555,7 +1555,7 @@ kOptimizerStatic8bit1StateBlockwise(T* p, T* __restrict__ const g, unsigned char
                 float* __restrict__ const quantiles1,
                 float* absmax1,
                 float weight_decay,
-                const float gnorm_scale, const int n)
+                const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
     //const int n_full = n + (n%BLOCK_SIZE);
@@ -1723,7 +1723,7 @@ MAKE_PreconditionOptimizer32bit1State(RMSPROP, float)
 
 #define MAKE_Optimizer32bit1State(oname, gtype) \
 template __global__ void kOptimizer32bit1State<gtype, oname>(gtype* g, gtype* p, float* state1, float *unorm, const float max_unorm, const float param_norm, \
-    const float beta1, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const int n); \
+    const float beta1, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n); \
 
 MAKE_Optimizer32bit1State(MOMENTUM, half)
 MAKE_Optimizer32bit1State(MOMENTUM, float)
@@ -1740,9 +1740,9 @@ MAKE_PreconditionOptimizer32bit2State(ADAM, half)
 MAKE_PreconditionOptimizer32bit2State(ADAM, float)
 
 template __global__ void kOptimizer32bit2State<half, ADAM>(half* g, half* p, float* state1, float* state2, float *unorm, const float max_unorm, const float param_norm,
-    const float beta1, const float beta2, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const int n);
+    const float beta1, const float beta2, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 template __global__ void kOptimizer32bit2State<float, ADAM>(float* g, float* p, float* state1, float* state2, float *unorm, const float max_unorm, const float param_norm,
-    const float beta1, const float beta2, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const int n);
+    const float beta1, const float beta2, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 #define MAKE_PreconditionStatic8bit1State(oname, gtype) \
 template __global__ void kPreconditionOptimizerStatic8bit1State<gtype, oname>(gtype* p, gtype* __restrict__ const g, unsigned char*__restrict__  const state1,  \
@@ -1825,7 +1825,7 @@ template __global__ void kOptimizerStatic8bit2StateBlockwise<gtype, oname, block
                 float* __restrict__ const quantiles1, float* __restrict__ const quantiles2, \
                 float* absmax1, float* absmax2,  \
                 float weight_decay, \
-                const float gnorm_scale, const int n); \
+                const float gnorm_scale, const bool skip_zeros, const int n); \
 
 MAKE_OptimizerStatic8bit2StateBlockwise(ADAM, float, 2048, 8)
 MAKE_OptimizerStatic8bit2StateBlockwise(ADAM, half, 2048, 8)
@@ -1838,7 +1838,7 @@ template __global__ void kOptimizerStatic8bit1StateBlockwise<gtype, oname, block
                 float* __restrict__ const quantiles1, \
                 float* absmax1, \
                 float weight_decay, \
-                const float gnorm_scale, const int n); \
+                const float gnorm_scale, const bool skip_zeros, const int n); \
 
 MAKE_OptimizerStatic8bit1StateBlockwise(MOMENTUM, float, 2048, 8)
 MAKE_OptimizerStatic8bit1StateBlockwise(MOMENTUM, half, 2048, 8)

csrc/kernels.cuh

@@ -27,7 +27,7 @@ template<typename T, int OPTIMIZER>
 __global__ void kOptimizer32bit2State(T* g, T* p, 
                 float* state1, float* state2, float *unorm, const float max_unorm, const float param_norm,
                 const float beta1, const float beta2, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, const int n);
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 template<typename T, int OPTIMIZER, int BLOCK_SIZE, int NUM_VALS>
 __global__ void kPreconditionOptimizer32bit1State(T* g, T* p, 
@@ -39,7 +39,7 @@ template<typename T, int OPTIMIZER>
 __global__ void kOptimizer32bit1State(T* g, T* p, 
                 float* state1,  float *unorm, const float max_unorm, const float param_norm,
                 const float beta1, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, const int n);
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 template<typename T, int OPTIMIZER>
 __global__ void
@@ -90,7 +90,7 @@ template<typename T, int OPTIMIZER, int BLOCK_SIZE, int N_PER_TH> __global__ voi
 		T* p, T* __restrict__ const g, unsigned char* state1, unsigned char* state2,
                 const float beta1, const float beta2, const float eps, const int step, const float lr,
                 float* __restrict__ const quantiles1, float* __restrict__ const quantiles2,
-                float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, const int n);
+                float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, const bool skip_zeros, const int n);
 
 template<typename T, int OPTIMIZER, int BLOCK_SIZE, int N_PER_TH> __global__ void kOptimizerStatic8bit1StateBlockwise(
 		T* p, T* __restrict__ const g, unsigned char* state1,
@@ -99,7 +99,7 @@ template<typename T, int OPTIMIZER, int BLOCK_SIZE, int N_PER_TH> __global__ voi
                 float* __restrict__ const quantiles1,
                 float* absmax1,
                 float weight_decay,
-                const float gnorm_scale, const int n);
+                const float gnorm_scale, const bool skip_zeros, const int n);
 
 
 template<typename T, int BLOCK_SIZE, int NUM_VALS> __global__ void kPercentileClipping(T * __restrict__ g, float *gnorm_vec, int step, const int n);

csrc/ops.cu

@@ -181,7 +181,7 @@ template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, flo
 template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p, 
                 float* state1, float* state2, float *unorm, float max_unorm, float param_norm,
                 const float beta1, const float beta2, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, const int n)
+                const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n)
 {
   int blocks = n/4096;
   blocks = n % 4096 == 0 ? blocks : blocks + 1;
@@ -194,7 +194,7 @@ template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p,
         kPreconditionOptimizer32bit2State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
         CUDA_CHECK_RETURN(cudaPeekAtLastError());
       }
-			kOptimizer32bit2State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
+			kOptimizer32bit2State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
       CUDA_CHECK_RETURN(cudaPeekAtLastError());
 			break;
 		case MOMENTUM:
@@ -206,7 +206,7 @@ template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p,
         CUDA_CHECK_RETURN(cudaPeekAtLastError());
 			}
 
-			kOptimizer32bit1State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, n);
+			kOptimizer32bit1State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
       CUDA_CHECK_RETURN(cudaPeekAtLastError());
 			break;
 	}
@@ -259,7 +259,7 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g,
 
 template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g,
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr, 
-                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, int n)
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)
 {
 
 	int blocks = 0;
@@ -269,7 +269,7 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g
 			blocks = n/BLOCKSIZE_2STATE;
 			blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1;
 			kOptimizerStatic8bit2StateBlockwise<T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE><<<blocks, BLOCKSIZE_2STATE/NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr,
-																														quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, n);
+																														quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n);
 			CUDA_CHECK_RETURN(cudaPeekAtLastError());
 		break;
 		case MOMENTUM:
@@ -277,7 +277,7 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g
 			blocks = n/BLOCKSIZE_1STATE;
 			blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1;
 			kOptimizerStatic8bit1StateBlockwise<T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE><<<blocks, BLOCKSIZE_1STATE/NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr,
-																														quantiles1, absmax1, weight_decay, gnorm_scale, n);
+																														quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n);
 			CUDA_CHECK_RETURN(cudaPeekAtLastError());
 		break;
 	}
@@ -313,7 +313,7 @@ template void dequantizeBlockwise<float>(float *code, unsigned char *A, float *a
 template void optimizer32bit<gtype, name>(gtype* g, gtype* p, \
                 float* state1, float* state2, float* unorm, float max_unorm, float param_norm, \
                 const float beta1, const float beta2, const float eps, const float weight_decay, \
-                const int step, const float lr, const float gnorm_scale, const int n);
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 MAKE_optimizer32bit(ADAM, half)
 MAKE_optimizer32bit(ADAM, float)
@@ -342,7 +342,7 @@ MAKE_optimizerStatic8bit(RMSPROP, float)
 #define MAKE_optimizerStatic8bitBlockwise(gtype, optim_name) \
 template void optimizerStatic8bitBlockwise<gtype, optim_name>(gtype* p, gtype* g, \
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,  \
-                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, int n); \
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n); \
 
 MAKE_optimizerStatic8bitBlockwise(half, ADAM);
 MAKE_optimizerStatic8bitBlockwise(float, ADAM);

csrc/ops.cuh

@@ -49,7 +49,7 @@ template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, flo
 template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p, 
                 float* state1, float* state2, float *unorm, float max_unorm, float param_norm,
                 float beta1, float beta2, float eps, float weight_decay,
-                int step, float lr, const float gnorm_scale, int n);
+                int step, float lr, const float gnorm_scale, bool skip_zeros, int n);
 
 template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g, unsigned char* state1, unsigned char* state2,
                 float *unorm, float max_unorm, float param_norm,
@@ -62,7 +62,8 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g, unsigne
 
 template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g,
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr, 
-                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, int n);
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, 
+								bool skip_zeros, int n);
 
 template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step, const int n);
 

csrc/pythonInterface.c

@@ -20,8 +20,8 @@ void estimateQuantiles_fp16(half *A, float *code, float offset, int n){ estimate
 void fname##32bit_g##gbits(gtype *g, gtype *p, \
                float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
                const float beta1, const float beta2, const float eps, const float weight_decay, \
-               const int step, const float lr, float gnorm_scale, const int n) \
+               const int step, const float lr, float gnorm_scale, bool skip_zeros, const int n) \
-{ optimizer32bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n); } \
+{ optimizer32bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
 
 MAKE_FUNC32(momentum, MOMENTUM, float, 32)
 MAKE_FUNC32(momentum, MOMENTUM, half, 16)
@@ -53,8 +53,8 @@ MAKE_FUNC8(rmsprop, RMSPROP, half, 16)
 #define MAKE_BLOCKWISE8(fname, optim_name, gtype, gbits) \
 void fname##_8bit_blockwise_fp##gbits(gtype* p, gtype* g, \
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr, \
-                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, int n)\
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)\
-{	optimizerStatic8bitBlockwise<gtype, optim_name>(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, n); }\
+{	optimizerStatic8bitBlockwise<gtype, optim_name>(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); }\
 
 MAKE_BLOCKWISE8(adam, ADAM, half, 16)
 MAKE_BLOCKWISE8(adam, ADAM, float, 32)
@@ -93,8 +93,8 @@ extern "C"
 	void c##name##32bit_g##gbits(gtype *g, gtype *p, \
 								 float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
 								 const float beta1, const float beta2, const float eps, const float weight_decay, \
-								 const int step, const float lr, const float gnorm_scale, const int n) \
+								 const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \
-	{ name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n); } \
+	{ name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
 
 	MAKE_CFUNC32(adam, float, 32)
 	MAKE_CFUNC32(adam, half, 16)
@@ -110,7 +110,7 @@ extern "C"
                 float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, \
                 float* max1, float* max2, float* new_max1, float* new_max2, \
-                float weight_decay, float gnorm_scale, int n) \
+                float weight_decay, float gnorm_scale, bool skip_zeros, int n) \
   {  \
 	    name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
 			                                 quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
@@ -126,8 +126,8 @@ extern "C"
   #define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \
   void c##fname##_8bit_blockwise_fp##gbits(gtype* p, gtype* g, \
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,  \
-                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, int n) \
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n) \
-  {	fname##_8bit_blockwise_fp##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, n); } \
+  {	fname##_8bit_blockwise_fp##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); } \
 
 	MAKE_CBLOCKWISE8(adam, ADAM, half, 16)
 	MAKE_CBLOCKWISE8(adam, ADAM, float, 32)

tests/test_optim.py

@@ -141,6 +141,7 @@ def test_global_config(dim1, dim2, gtype):
     eps = 1e-8
 
     bnb.optim.GlobalOptimManager.get_instance().initialize()
+    bnb.optim.GlobalOptimManager.get_instance().override_config(p2, 'skip_zeros', True)
     bnb.optim.GlobalOptimManager.get_instance().override_config(p3, 'optim_bits', 8)
 
     bnb.optim.GlobalOptimManager.get_instance().register_parameters([p1, p2, p3])
@@ -155,6 +156,8 @@ def test_global_config(dim1, dim2, gtype):
     else:
         atol, rtol = 1e-4, 1e-3
 
+    original_p2 = p2[mask].clone()
+
     for i in range(50):
         g1 = torch.randn(dim1,dim2, device='cuda', dtype=gtype)*0.1 + 0.001
         g2 = torch.randn(dim1,dim2, device='cuda', dtype=gtype)*0.1 + 0.001
@@ -163,11 +166,32 @@ def test_global_config(dim1, dim2, gtype):
         p2.grad = g2
         p3.grad = g3
 
+        if i > 30 and i % 10 == 0:
+            g1.data[mask] = 0.0
+            g2.data[mask] = 0.0
+            p1.grad = g1
+            p2.grad = g2
+            original_p1 = p1[mask].clone()
+            original_p2 = p2[mask].clone()
+            og_s1 = adam2.state[p2]['state1'][mask].clone()
+            og_s2 = adam2.state[p2]['state2'][mask].clone()
+            og_s11 = adam2.state[p1]['state1'][mask].clone()
+            og_s21 = adam2.state[p1]['state2'][mask].clone()
+
         adam2.step()
 
         assert adam2.state[p3]['state1'].dtype == torch.uint8
         assert adam2.state[p3]['state2'].dtype == torch.uint8
 
+        if i > 30 and i % 10 == 0:
+            torch.testing.assert_allclose(original_p2, p2[mask])
+            torch.testing.assert_allclose(adam2.state[p2]['state1'][mask], og_s1)
+            torch.testing.assert_allclose(adam2.state[p2]['state2'][mask], og_s2)
+            assert ((p1[mask]- original_p1)==0.0).sum() < p1.numel()
+            assert ((adam2.state[p1]['state1'][mask]- og_s11)==0.0).sum() == 0.0
+            assert ((adam2.state[p1]['state2'][mask]- og_s21)==0.0).sum() == 0.0
+
+
 
 
 dim1 = [1024]