diff --git a/test/mockgpu/nv/nvgpu.py b/test/mockgpu/nv/nvgpu.py
index 8a7f8e3074..f450a16ab1 100644
--- a/test/mockgpu/nv/nvgpu.py
+++ b/test/mockgpu/nv/nvgpu.py
@@ -92,7 +92,7 @@ class GPFIFO:
     qmd = qmd_struct_t.from_address(qmd_addr)
     prg_addr = qmd.program_address_lower + (qmd.program_address_upper << 32)
     const0 = to_mv(qmd.constant_buffer_addr_lower_0 + (qmd.constant_buffer_addr_upper_0 << 32), 0x160).cast('I')
-    args_cnt, vals_cnt = const0[0], const0[1]
+    args_cnt, vals_cnt = const0[80], const0[81]
     args_addr = qmd.constant_buffer_addr_lower_0 + (qmd.constant_buffer_addr_upper_0 << 32) + 0x160
     args = to_mv(args_addr, args_cnt*8).cast('Q')
     vals = to_mv(args_addr + args_cnt*8, vals_cnt*4).cast('I')
diff --git a/tinygrad/runtime/ops_nv.py b/tinygrad/runtime/ops_nv.py
index 368b3c03b9..679823e69e 100644
--- a/tinygrad/runtime/ops_nv.py
+++ b/tinygrad/runtime/ops_nv.py
@@ -141,6 +141,7 @@ class NVComputeQueue(NVCommandQueue):
 
     self.bind_sints_to_ptr(*global_size, ptr=qmd_addr + nv_gpu.NVC6C0_QMDV03_00_CTA_RASTER_WIDTH[1] // 8, fmt='I')
     self.bind_sints_to_ptr(*local_size, ptr=qmd_addr + nv_gpu.NVC6C0_QMDV03_00_CTA_THREAD_DIMENSION0[1] // 8, fmt='H')
+    self.bind_sints_to_ptr(*local_size, *global_size, ptr=args_state.ptr, fmt='I')
     qmd.constant_buffer_addr_upper_0, qmd.constant_buffer_addr_lower_0 = data64(args_state.ptr)
 
     if self.active_qmd is None:
@@ -188,7 +189,7 @@ class NVCopyQueue(NVCommandQueue):
 
 class NVArgsState(CLikeArgsState):
   def __init__(self, ptr:int, prg:NVProgram, bufs:tuple[HCQBuffer, ...], vals:tuple[int, ...]=()):
-    if MOCKGPU: prg.constbuffer_0[0:2] = [len(bufs), len(vals)]
+    if MOCKGPU: prg.constbuffer_0[80:82] = [len(bufs), len(vals)]
     super().__init__(ptr, prg, bufs, vals=vals, prefix=prg.constbuffer_0)
 
 class NVProgram(HCQProgram):