[MOE] Standardize and abstract fused shared expert optimization selection (#51695)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
fxmarty-amd
2026-08-18 12:56:56 -04:00
committed by GitHub
parent f9f066d195
commit 88b2bff2c6
20 changed files with 1288 additions and 281 deletions
@@ -576,63 +576,91 @@ class QuarkConfig(QuantizationConfig):
return True
def _find_matched_config(
self, layer_name: str, module: torch.nn.Module
) -> dict[str, Any]:
def get_layer_quant_config_from_name(
self, layer_name: str
) -> dict[str, Any] | None:
proj_name = layer_name.split(".")[-1]
if proj_name in self.packed_modules_mapping:
shard_proj_names = self.packed_modules_mapping[proj_name]
# Convert fused_name --> [shard_names]
shard_names = [
layer_name.replace(proj_name, shard_proj_name)
for shard_proj_name in shard_proj_names
]
shard_configs = []
for shard_name in shard_names:
if shard_name == layer_name:
config = cast(
dict[str, Any], self.quant_config.get("global_quant_config")
)
for shard_proj_name in shard_proj_names:
shard_name = layer_name.replace(proj_name, shard_proj_name)
if shard_name != layer_name:
config = self.get_layer_quant_config_from_name(shard_name)
else:
config = self._find_matched_config(shard_name, module)
config = None
shard_configs.append(config)
if not all(
deep_compare(q_config, shard_configs[0]) for q_config in shard_configs
matched_configs = [config for config in shard_configs if config is not None]
if matched_configs and not all(
deep_compare(config, matched_configs[0]) for config in matched_configs
):
raise ValueError(
f"Found a different quantization configuration for "
f"{shard_proj_names} in {layer_name}. vLLM "
"requires all to use the same scheme."
)
return shard_configs[0]
if matched_configs:
return matched_configs[0]
return None
else:
layer_quant_config = cast(
dict[str, Any], self.quant_config.get("layer_quant_config")
dict[str, Any], self.quant_config.get("layer_quant_config") or {}
)
def _matches_pattern(layer_name, pattern):
if "*" not in pattern:
return layer_name in pattern
return fnmatch.fnmatch(layer_name, pattern)
for name_pattern, config in layer_quant_config.items():
if _matches_pattern(layer_name, name_pattern):
if "*" not in name_pattern:
matches = layer_name in name_pattern
else:
matches = fnmatch.fnmatch(layer_name, name_pattern)
if matches:
return config
return None
layer_type = cast(str, type(module))
layer_type_quant_config = cast(
dict[str, Any], self.quant_config.get("layer_type_quant_config")
)
if layer_type in layer_type_quant_config:
return layer_type_quant_config[layer_type]
def _find_matched_config(
self, layer_name: str, module: torch.nn.Module
) -> dict[str, Any]:
# Priority order:
# 1. layer_quant_config,
# 2. layer_type_quant_config,
# 3. global_quant_config.
global_quant_config = cast(
dict[str, Any], self.quant_config.get("global_quant_config")
)
return global_quant_config
layer_type = cast(str, type(module))
layer_type_quant_config = cast(
dict[str, Any], self.quant_config.get("layer_type_quant_config")
)
global_quant_config = cast(
dict[str, Any], self.quant_config.get("global_quant_config")
)
fallback_config = layer_type_quant_config.get(layer_type, global_quant_config)
proj_name = layer_name.split(".")[-1]
if proj_name in self.packed_modules_mapping:
shard_proj_names = self.packed_modules_mapping[proj_name]
shard_configs = []
for shard_proj_name in shard_proj_names:
shard_name = layer_name.replace(proj_name, shard_proj_name)
if shard_name == layer_name:
config = fallback_config
else:
config = self.get_layer_quant_config_from_name(shard_name)
if config is None:
config = fallback_config
shard_configs.append(config)
if not all(
deep_compare(config, shard_configs[0]) for config in shard_configs
):
raise ValueError(
f"Found a different quantization configuration for "
f"{shard_proj_names} in {layer_name}. vLLM requires all "
"to use the same scheme."
)
return shard_configs[0]
else:
layer_quant_config = self.get_layer_quant_config_from_name(layer_name)
if layer_quant_config is not None:
return layer_quant_config
return fallback_config
def _get_scheme_from_config(
self, config: dict[str, Any], dynamic_mxfp4_quant: bool = False