[Kernel/Quant] Remove the original marlin format and qqq (#23204)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-08-20 15:13:36 -04:00
committed by GitHub
parent ebe56a0064
commit 0cdbf5e61c
26 changed files with 92 additions and 3698 deletions
-36
View File
@@ -387,14 +387,6 @@ def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
# marlin
def marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
b_scales: torch.Tensor, workspace: torch.Tensor, size_m: int,
size_n: int, size_k: int) -> torch.Tensor:
return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
size_n, size_k)
# marlin_24
def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
b_meta: torch.Tensor, b_scales: torch.Tensor,
@@ -437,25 +429,6 @@ if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
is_zp_float: bool = False) -> torch.Tensor:
return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)
@register_fake("_C::marlin_qqq_gemm")
def _marlin_qqq_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
s_tok: torch.Tensor, s_ch: torch.Tensor,
s_group: torch.Tensor, workspace: torch.Tensor,
size_m: torch.SymInt, size_n: torch.SymInt,
size_k: torch.SymInt) -> torch.Tensor:
return torch.empty((size_m, size_n),
dtype=torch.float16,
device=a.device)
@register_fake("_C::marlin_gemm")
def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
b_scales: torch.Tensor, workspace: torch.Tensor,
size_m: torch.SymInt, size_n: torch.SymInt,
size_k: torch.SymInt) -> torch.Tensor:
return torch.empty((size_m, size_n),
dtype=torch.float16,
device=a.device)
@register_fake("_C::awq_dequantize")
def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
zeros: torch.Tensor, split_k_iters: torch.SymInt,
@@ -1348,15 +1321,6 @@ def scaled_int8_quant(
return output, input_scales, input_azp
# qqq ops
def marlin_qqq_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
s_tok: torch.Tensor, s_ch: torch.Tensor,
s_group: torch.Tensor, workspace: torch.Tensor,
size_m: int, size_n: int, size_k: int) -> torch.Tensor:
return torch.ops._C.marlin_qqq_gemm(a, b_q_weight, s_tok, s_ch, s_group,
workspace, size_m, size_n, size_k)
# gguf
def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int, n: int,
dtype: Optional[torch.dtype]) -> torch.Tensor: