Honor adam_w_mode in the CPU multi_tensor_adam binding

The CPU fused_adam extension created its Adam_Optimizer once with
default arguments and ignored the mode parameter, so FusedAdam on the
CPU backend always applied decoupled (AdamW) weight decay even when
constructed with adam_w_mode=False. Keep one optimizer instance per
mode instead; everything else is already passed per call.

This surfaced as cpu-torch-latest failures in the fp32-adam case of
test_fused_adam_matches_torch. The test's bf16 cases are dropped:
torch.optim in bf16 does its math in bf16 while the fused kernels
compute in fp32, so it was never a valid bf16 reference. Low-precision
dtypes get an explicit fp32-math reference in the FusedAdam rework.

Signed-off-by: PKUWZP <zhipeng.rainbowserie@gmail.com>

Author: PKUWZP

csrc/cpu/adam/fused_adam.cpp

@@ -19,13 +19,16 @@ void multi_tensor_adam(int chunk_size,
                        const int bias_correction,
                        const float weight_decay)
 {
-    static bool initialized = false;
-    if (!initialized) {
-        create_adam_optimizer(0);
-        initialized = true;
+    // ds_adam_step reads lr/betas/eps/weight_decay per call; only AdamW-vs-L2 is fixed at
+    // construction, so keep one optimizer instance per mode (mode 1 == AdamW, as in CUDA).
+    static bool initialized[2] = {false, false};
+    const int optimizer_id = mode;
+    if (!initialized[mode]) {
+        create_adam_optimizer(optimizer_id, 1e-3f, 0.9f, 0.999f, 1e-8f, 0.0f, mode == 1);
+        initialized[mode] = true;
     }
     for (int i = 0; i < tensor_lists[0].size(); i++) {
-        ds_adam_step(0,
+        ds_adam_step(optimizer_id,
                      step,
                      lr,
                      beta1,

tests/unit/ops/adam/test_adamw.py

@@ -81,7 +81,10 @@ def test(self,
 
 
 @pytest.mark.parametrize('adam_w_mode', [True, False], ids=["adamw", "adam"])
-@pytest.mark.parametrize('dtype', [torch.float, torch.bfloat16], ids=["fp32", "bf16"])
+# fp32 only: torch.optim in bf16 does its math in bf16 while the fused kernels compute in fp32,
+# so it is not a valid bf16 reference. Low-precision dtypes are covered against an explicit
+# fp32-math reference in the FusedAdam rework (#8300).
+@pytest.mark.parametrize('dtype', [torch.float], ids=["fp32"])
 def test_fused_adam_matches_torch(adam_w_mode, dtype):
     if dtype not in get_accelerator().supported_dtypes():
         pytest.skip(f"{dtype} not supported on {get_accelerator().device_name()}")
@@ -105,7 +108,6 @@ def test_fused_adam_matches_torch(adam_w_mode, dtype):
         ref_optimizer.step()
         ds_optimizer.step()
 
-    # bf16 storage rounds differently depending on where intermediates are kept, so allow one ulp.
-    atol = 1e-5 if dtype == torch.float else 2e-2
+    atol = 1e-5
     for ref_param, ds_param in zip(ref_params, ds_params):
         torch.testing.assert_close(ds_param.float(), ref_param.float(), atol=atol, rtol=0)