Support DeepSeek-V4 AMD Quark NVFP4 with emulation kernel (#47972)

Signed-off-by: Jimmy Adams <Adam_ji@outlook.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: huji <huidong.ji@amd.com>
Signed-off-by: jimmy-adams <41593649+jimmy-adams@users.noreply.github.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: huji <huidong.ji@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
jimmy-adams
2026-08-07 01:19:31 -07:00
committed by GitHub
co-authored by Andreas Karatzas huji Cursor
parent 0de0362ea1
commit da788334bc
11 changed files with 332 additions and 17 deletions
+17 -4
View File
@@ -750,6 +750,12 @@ class DeepseekV4Model(nn.Module, EagleModelMixin):
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
def _resolve_param_name(name: str) -> str:
inv_name = f"{name}_inv"
if name not in params_dict and inv_name in params_dict:
return inv_name
return name
# TP for attention
tp_size = get_tensor_model_parallel_world_size()
tp_rank = get_tensor_model_parallel_rank()
@@ -803,10 +809,11 @@ class DeepseekV4Model(nn.Module, EagleModelMixin):
if is_pp_missing_parameter(name, self):
break
param = params_dict[name]
param_name = _resolve_param_name(name)
param = params_dict[param_name]
weight_loader = param.weight_loader
weight_loader(param, loaded_weight, shard_id)
loaded_params.add(name)
loaded_params.add(param_name)
break
else:
if ".experts." in name:
@@ -857,12 +864,13 @@ class DeepseekV4Model(nn.Module, EagleModelMixin):
else:
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
param_name = _resolve_param_name(name)
param = params_dict[param_name]
weight_loader = getattr(
param, "weight_loader", default_weight_loader
)
weight_loader(param, loaded_weight)
loaded_params.add(name)
loaded_params.add(param_name)
continue
return loaded_params
@@ -934,6 +942,7 @@ def _make_deepseek_v4_weights_mapper(
"head.weight": "lm_head.weight",
"embed.weight": "embed_tokens.weight",
".ffn.gate.bias": ".ffn.gate.e_score_correction_bias",
".input_scale": ".input_scale_2",
},
orig_to_new_substr=substr_map,
)
@@ -945,6 +954,10 @@ class DeepseekV4ForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
# Default mapper assumes the original FP4-expert checkpoint layout.
# Overridden per-instance in __init__ when expert_dtype != "fp4".
hf_to_vllm_mapper = _make_deepseek_v4_weights_mapper("fp4")
packed_modules_mapping = {
"gate_up_proj": ["w1", "w3"],
"fused_wqa_wkv": ["wq_a", "wkv"],
}
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()