change quantization calibration method

Author: susanbao

src/maxdiffusion/configs/base_wan_14b.yml

@@ -319,8 +319,10 @@ quantization: ''
 quantization_local_shard_count: -1
 compile_topology_num_slices: -1 # Number of target slices, set to a positive integer.
 use_qwix_quantization: False # Whether to use qwix for quantization. If set to True, the transformer of WAN will be quantized using qwix.
-# Quantization calibration method used for weights and activations. Supported methods can be found in https://github.com/google/qwix/blob/dc2a0770351c740e5ab3cce7c0efe9f7beacce9e/qwix/qconfig.py#L70-L80
-quantization_calibration_method: "absmax"
+# Quantization calibration method used for weights, activations and bwd. Supported methods can be found in https://github.com/google/qwix/blob/dc2a0770351c740e5ab3cce7c0efe9f7beacce9e/qwix/qconfig.py#L70-L80
+weight_quantization_calibration_method: "absmax"
+act_quantization_calibration_method: "absmax"
+bwd_quantization_calibration_method: "absmax"
 qwix_module_path: ".*"
 
 # Eval model on per eval_every steps. -1 means don't eval.

src/maxdiffusion/pipelines/wan/wan_pipeline.py

@@ -302,9 +302,9 @@ def get_fp8_config(cls, config: HyperParameters):
             act_qtype=jnp.float8_e4m3fn,
             bwd_qtype=jnp.float8_e5m2,
             disable_channelwise_axes=True,  # per_tensor calibration
-            weight_calibration_method=config.quantization_calibration_method,
-            act_calibration_method=config.quantization_calibration_method,
-            bwd_calibration_method=config.quantization_calibration_method,
+            weight_calibration_method=config.weight_quantization_calibration_method,
+            act_calibration_method=config.act_quantization_calibration_method,
+            bwd_calibration_method=config.bwd_quantization_calibration_method,
             op_names=("dot_general", "einsum"),
         ),
         qwix.QtRule(
@@ -313,9 +313,9 @@ def get_fp8_config(cls, config: HyperParameters):
             act_qtype=jnp.float8_e4m3fn,
             bwd_qtype=jnp.float8_e4m3fn,
             disable_channelwise_axes=True,  # per_tensor calibration
-            weight_calibration_method=config.quantization_calibration_method,
-            act_calibration_method=config.quantization_calibration_method,
-            bwd_calibration_method=config.quantization_calibration_method,
+            weight_calibration_method=config.weight_quantization_calibration_method,
+            act_calibration_method=config.act_quantization_calibration_method,
+            bwd_calibration_method=config.bwd_quantization_calibration_method,
             op_names=("conv_general_dilated"),
         ),
     ]