sglang 0.4.7__py3-none-any.whl → 0.4.7.post1__py3-none-any.whl

sglang/srt/layers/pooler.py

@@ -3,10 +3,13 @@
 
 from dataclasses import dataclass
 from enum import IntEnum
+from typing import Optional
 
 import torch
 import torch.nn as nn
+from transformers import PretrainedConfig
 
+from sglang.srt.layers.activation import get_cross_encoder_activation_function
 from sglang.srt.model_executor.model_runner import ForwardBatch
 
 
@@ -54,3 +57,56 @@ class Pooler(nn.Module):
             pooled_data = nn.functional.normalize(pooled_data, p=2, dim=1)
 
         return EmbeddingPoolerOutput(embeddings=pooled_data)
+
+
+class CrossEncodingPooler(nn.Module):
+    """A layer that pools specific information from hidden states.
+
+    This layer does the following:
+    1. Extracts specific tokens or aggregates data based on pooling method.
+    2. Normalizes output if specified.
+    3. Returns structured results as `EmbeddingPoolerOutput`.
+    """
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        classifier: nn.Module,
+        pooler: Optional[nn.Module] = None,
+    ):
+        super().__init__()
+        self.classifier = classifier
+        self.pooler = pooler
+        self.default_activation_function = get_cross_encoder_activation_function(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> EmbeddingPoolerOutput:
+        """Pools sentence pair scores from the hidden_states."""
+
+        prompt_lens = forward_batch.extend_seq_lens
+
+        offset = 0
+        pooled_data_lst = []
+        for prompt_len in prompt_lens:
+            pooled_data_i = hidden_states[offset : offset + prompt_len]
+
+            if self.pooler is not None:
+                final_shape_tensor = self.pooler(pooled_data_i, forward_batch)
+            else:
+                final_shape_tensor = self.classifier(pooled_data_i)
+
+            pooled_data_lst.append(final_shape_tensor)
+            offset += prompt_len
+
+        pooled_output = torch.stack(pooled_data_lst)
+
+        if self.pooler is not None:
+            # apply classifier once on the full batch if possible
+            pooled_output = self.classifier(pooled_output)
+
+        scores = self.default_activation_function(pooled_output).squeeze(-1)
+
+        return EmbeddingPoolerOutput(embeddings=scores)

sglang/srt/layers/quantization/deep_gemm_wrapper/__init__.py

@@ -0,0 +1 @@
+from .entrypoint import *

sglang/srt/layers/quantization/{deep_gemm.py → deep_gemm_wrapper/compile_utils.py}

@@ -5,34 +5,23 @@ from dataclasses import dataclass
 from enum import IntEnum, auto
 from typing import Callable, Dict, List, Optional, Tuple
 
-import torch
 from tqdm.contrib.concurrent import thread_map
 
+from sglang.srt.layers.quantization.deep_gemm_wrapper.configurer import (
+    DEEPGEMM_BLACKWELL,
+    ENABLE_JIT_DEEPGEMM,
+)
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import get_bool_env_var, get_device_sm, get_int_env_var, is_cuda
+from sglang.srt.utils import get_bool_env_var, get_int_env_var
 
 logger = logging.getLogger(__name__)
-_ENABLE_JIT_DEEPGEMM = False
 
-try:
-    import deep_gemm
+if ENABLE_JIT_DEEPGEMM and not DEEPGEMM_BLACKWELL:
     from deep_gemm import get_num_sms
     from deep_gemm.jit import build
-    from deep_gemm.jit.compiler import get_nvcc_compiler
     from deep_gemm.jit_kernels.gemm import get_best_configs
     from deep_gemm.jit_kernels.runtime import FP8GemmRuntime, GemmType
 
-    sm_version = get_device_sm()
-    if sm_version == 90:
-        if get_bool_env_var("SGL_ENABLE_JIT_DEEPGEMM", default="true"):
-            _ENABLE_JIT_DEEPGEMM = True
-except ImportError:
-    logger.warning("Failed to import deepgemm, disable _ENABLE_JIT_DEEPGEMM.")
-
-
-def get_enable_jit_deepgemm():
-    return _ENABLE_JIT_DEEPGEMM
-
 
 _BUILTIN_M_LIST = list(range(1, 1024 * 16 + 1))
 _ENABLE_JIT_DEEPGEMM_PRECOMPILE = get_bool_env_var(
@@ -52,8 +41,10 @@ os.environ["DG_JIT_CACHE_DIR"] = os.getenv(
 # NVRTC may have performance loss with some cases.
 # And NVCC JIT speed is also 9x faster in the ref commit
 _USE_NVRTC_DEFAULT = "0"
-if _ENABLE_JIT_DEEPGEMM:
+if ENABLE_JIT_DEEPGEMM:
     try:
+        from deep_gemm.jit.compiler import get_nvcc_compiler
+
         get_nvcc_compiler()
     except:
         logger.warning(
@@ -114,11 +105,12 @@ class DeepGemmKernelHelper:
 _INITIALIZATION_DICT: Dict[Tuple[DeepGemmKernelType, int, int, int], bool] = dict()
 
 
+# TODO improve naming
 def _compile_warning_1():
     if not _IN_PRECOMPILE_STAGE and _IS_FIRST_RANK_ON_NODE:
         logger.warning(
             "Entering DeepGEMM JIT Pre-Compile session. "
-            "And it may takes a long time(Typically 10-20 mins) "
+            "It may takes a long time (typically 10-20 mins) "
             "if you have not run `sglang.compile_deep_gemm`. "
             "It is recommended to run `sglang.compile_deep_gemm` with same args as `sglang.launch_server`"
             " for pre-compilation to reduce the overhead if you have not run it before. "
@@ -127,6 +119,7 @@ def _compile_warning_1():
         )
 
 
+# TODO improve naming
 def _compile_warning_2():
     logger.warning(
         "Entering DeepGEMM JIT Single Kernel Compile session. "
@@ -238,6 +231,7 @@ def _compile_gemm_nt_f8f8bf16_one(
     _ = build("gemm_fp8_fp8_bf16_nt", code, FP8GemmRuntime, kwargs)
 
 
+# TODO further refactor warmup-related
 _KERNEL_HELPER_DICT: Dict[DeepGemmKernelType, DeepGemmKernelHelper] = {
     DeepGemmKernelType.GROUPED_GEMM_NT_F8F8BF16_MASKED: DeepGemmKernelHelper(
         name="m_grouped_gemm_fp8_fp8_bf16_nt_masked",
@@ -270,7 +264,6 @@ def _maybe_compile_deep_gemm_one_type_all(
     num_groups: int,
     m_list: Optional[List[int]] = None,
 ) -> None:
-
     global _INITIALIZATION_DICT
     global _BUILTIN_M_LIST
 
@@ -304,56 +297,6 @@ def _maybe_compile_deep_gemm_one_type_all(
         thread_map(compile_func, collected_configs, max_workers=_COMPILE_WORKERS)
 
 
-def grouped_gemm_nt_f8f8bf16_masked(
-    lhs: Tuple[torch.Tensor, torch.Tensor],
-    rhs: Tuple[torch.Tensor, torch.Tensor],
-    out: torch.Tensor,
-    masked_m: torch.Tensor,
-    expected_m: int,
-):
-    num_groups, _, k = lhs[0].shape
-    _, n, _ = rhs[0].shape
-
-    kernel_type = DeepGemmKernelType.GROUPED_GEMM_NT_F8F8BF16_MASKED
-    _maybe_compile_deep_gemm_one_type_all(kernel_type, n, k, num_groups)
-
-    with _log_jit_build(expected_m, n, k, kernel_type):
-        deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(
-            lhs, rhs, out, masked_m, expected_m
-        )
-
-
-def grouped_gemm_nt_f8f8bf16_contig(
-    lhs: Tuple[torch.Tensor, torch.Tensor],
-    rhs: Tuple[torch.Tensor, torch.Tensor],
-    out: torch.Tensor,
-    m_indices: torch.Tensor,
-):
-    m, k = lhs[0].shape
-    num_groups, n, _ = rhs[0].shape
-
-    kernel_type = DeepGemmKernelType.GROUPED_GEMM_NT_F8F8BF16_CONTIG
-    _maybe_compile_deep_gemm_one_type_all(kernel_type, n, k, num_groups)
-
-    with _log_jit_build(m, n, k, kernel_type):
-        deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(lhs, rhs, out, m_indices)
-
-
-def gemm_nt_f8f8bf16(
-    lhs: Tuple[torch.Tensor, torch.Tensor],
-    rhs: Tuple[torch.Tensor, torch.Tensor],
-    out: torch.Tensor,
-):
-    m, k = lhs[0].shape
-    n, _ = rhs[0].shape
-
-    kernel_type = DeepGemmKernelType.GEMM_NT_F8F8BF16
-    _maybe_compile_deep_gemm_one_type_all(kernel_type, n, k, 1)
-
-    with _log_jit_build(m, n, k, kernel_type):
-        deep_gemm.gemm_fp8_fp8_bf16_nt(lhs, rhs, out)
-
-
 @contextmanager
 def _log_jit_build(M: int, N: int, K: int, kernel_type: DeepGemmKernelType):
     if _IN_PRECOMPILE_STAGE:
@@ -368,7 +311,8 @@ def _log_jit_build(M: int, N: int, K: int, kernel_type: DeepGemmKernelType):
         ret = origin_func(self, *args, **kwargs)
         if ret is None:
             kernel_helper = _KERNEL_HELPER_DICT[kernel_type]
-            _compile_warning_2()
+            if not DEEPGEMM_BLACKWELL:
+                _compile_warning_2()
             logger.warning(
                 f"DeepGEMM JIT Compiling for <{kernel_helper.name}> M={M}, N={N}, K={K}. Please wait."
             )
@@ -380,13 +324,12 @@ def _log_jit_build(M: int, N: int, K: int, kernel_type: DeepGemmKernelType):
 
 
 @contextmanager
-def configure_deep_gemm_num_sms(num_sms):
-    if num_sms is None:
+def deep_gemm_execution_hook(
+    m: int, n: int, k: int, num_groups: int, kernel_type: DeepGemmKernelType
+):
+    # not supported yet
+    if not DEEPGEMM_BLACKWELL:
+        _maybe_compile_deep_gemm_one_type_all(kernel_type, n, k, num_groups)
+
+    with _log_jit_build(m, n, k, kernel_type):
         yield
-    else:
-        original_num_sms = deep_gemm.get_num_sms()
-        deep_gemm.set_num_sms(num_sms)
-        try:
-            yield
-        finally:
-            deep_gemm.set_num_sms(original_num_sms)

sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py

@@ -0,0 +1,32 @@
+import logging
+
+from sglang.srt.utils import get_bool_env_var, get_device_sm
+
+logger = logging.getLogger(__name__)
+
+
+def _compute_enable_deep_gemm():
+    sm_version = get_device_sm()
+    if sm_version < 90:
+        return False
+
+    try:
+        import deep_gemm
+    except ImportError:
+        logger.warning("Failed to import deep_gemm, disable ENABLE_JIT_DEEPGEMM.")
+        return False
+
+    return get_bool_env_var("SGL_ENABLE_JIT_DEEPGEMM", default="true")
+
+
+ENABLE_JIT_DEEPGEMM = _compute_enable_deep_gemm()
+
+try:
+    from deep_gemm import fp8_gemm_nt
+
+    # They have not given a name to this breaking change
+    DEEPGEMM_BLACKWELL = True
+except ImportError:
+    DEEPGEMM_BLACKWELL = False
+
+DEEPGEMM_SCALE_UE8M0 = DEEPGEMM_BLACKWELL

sglang/srt/layers/quantization/deep_gemm_wrapper/entrypoint.py

@@ -0,0 +1,110 @@
+import logging
+from contextlib import contextmanager
+from typing import Tuple
+
+import torch
+
+from sglang.srt.layers.quantization.deep_gemm_wrapper import compile_utils
+from sglang.srt.layers.quantization.deep_gemm_wrapper.configurer import (
+    DEEPGEMM_BLACKWELL,
+    DEEPGEMM_SCALE_UE8M0,
+    ENABLE_JIT_DEEPGEMM,
+)
+from sglang.srt.server_args import ServerArgs
+
+logger = logging.getLogger(__name__)
+
+if ENABLE_JIT_DEEPGEMM:
+    import deep_gemm
+
+    if DEEPGEMM_BLACKWELL:
+        from deep_gemm import fp8_gemm_nt as _gemm_nt_f8f8bf16_raw
+        from deep_gemm import (
+            fp8_m_grouped_gemm_nt_masked as _grouped_gemm_nt_f8f8bf16_masked_raw,
+        )
+        from deep_gemm import (
+            m_grouped_fp8_gemm_nt_contiguous as _grouped_gemm_nt_f8f8bf16_contig_raw,
+        )
+    else:
+        from deep_gemm import gemm_fp8_fp8_bf16_nt as _gemm_nt_f8f8bf16_raw
+        from deep_gemm import get_col_major_tma_aligned_tensor
+        from deep_gemm import (
+            m_grouped_gemm_fp8_fp8_bf16_nt_contiguous as _grouped_gemm_nt_f8f8bf16_contig_raw,
+        )
+        from deep_gemm import (
+            m_grouped_gemm_fp8_fp8_bf16_nt_masked as _grouped_gemm_nt_f8f8bf16_masked_raw,
+        )
+
+
+def grouped_gemm_nt_f8f8bf16_masked(
+    lhs: Tuple[torch.Tensor, torch.Tensor],
+    rhs: Tuple[torch.Tensor, torch.Tensor],
+    out: torch.Tensor,
+    masked_m: torch.Tensor,
+    expected_m: int,
+    recipe=None,
+):
+    num_groups, _, k = lhs[0].shape
+    _, n, _ = rhs[0].shape
+    kernel_type = compile_utils.DeepGemmKernelType.GROUPED_GEMM_NT_F8F8BF16_MASKED
+
+    with compile_utils.deep_gemm_execution_hook(
+        expected_m, n, k, num_groups, kernel_type
+    ):
+        _grouped_gemm_nt_f8f8bf16_masked_raw(
+            lhs,
+            rhs,
+            out,
+            masked_m,
+            expected_m,
+            **({"recipe": recipe} if DEEPGEMM_BLACKWELL else {})
+        )
+
+
+def grouped_gemm_nt_f8f8bf16_contig(
+    lhs: Tuple[torch.Tensor, torch.Tensor],
+    rhs: Tuple[torch.Tensor, torch.Tensor],
+    out: torch.Tensor,
+    m_indices: torch.Tensor,
+):
+    m, k = lhs[0].shape
+    num_groups, n, _ = rhs[0].shape
+    kernel_type = compile_utils.DeepGemmKernelType.GROUPED_GEMM_NT_F8F8BF16_CONTIG
+
+    with compile_utils.deep_gemm_execution_hook(m, n, k, num_groups, kernel_type):
+        _grouped_gemm_nt_f8f8bf16_contig_raw(lhs, rhs, out, m_indices)
+
+
+def gemm_nt_f8f8bf16(
+    lhs: Tuple[torch.Tensor, torch.Tensor],
+    rhs: Tuple[torch.Tensor, torch.Tensor],
+    out: torch.Tensor,
+):
+    m, k = lhs[0].shape
+    n, _ = rhs[0].shape
+    num_groups = 1
+    kernel_type = compile_utils.DeepGemmKernelType.GEMM_NT_F8F8BF16
+
+    with compile_utils.deep_gemm_execution_hook(m, n, k, num_groups, kernel_type):
+        _gemm_nt_f8f8bf16_raw(
+            lhs,
+            rhs,
+            out,
+        )
+
+
+def update_deep_gemm_config(gpu_id: int, server_args: ServerArgs):
+    compile_utils.update_deep_gemm_config(gpu_id, server_args)
+
+
+@contextmanager
+def configure_deep_gemm_num_sms(num_sms):
+    if num_sms is None:
+        yield
+    else:
+        original_num_sms = deep_gemm.get_num_sms()
+        deep_gemm.set_num_sms(num_sms)
+        try:
+            yield
+        finally:
+            deep_gemm.set_num_sms(original_num_sms)

sglang/srt/layers/quantization/fp8_kernel.py

@@ -23,7 +23,8 @@ import torch
 import triton
 import triton.language as tl
 
-from sglang.srt.layers.quantization.deep_gemm import _ENABLE_JIT_DEEPGEMM
+from sglang.math_utils import align
+from sglang.srt.layers.quantization import deep_gemm_wrapper
 from sglang.srt.utils import (
     direct_register_custom_op,
     get_device_core_count,
@@ -44,10 +45,6 @@ if _is_cuda:
         sgl_per_token_quant_fp8,
     )
 
-    from sglang.srt.layers.quantization.deep_gemm import (
-        gemm_nt_f8f8bf16 as deep_gemm_gemm_nt_f8f8bf16,
-    )
-
 logger = logging.getLogger(__name__)
 
 
@@ -67,7 +64,6 @@ else:
     fp8_max = torch.finfo(fp8_dtype).max
 fp8_min = -fp8_max
 
-
 if supports_custom_op():
 
     def deep_gemm_fp8_fp8_bf16_nt(
@@ -77,7 +73,7 @@ if supports_custom_op():
         Bs: torch.Tensor,
         C: torch.Tensor,
     ) -> None:
-        deep_gemm_gemm_nt_f8f8bf16((A, As), (B, Bs), C)
+        deep_gemm_wrapper.gemm_nt_f8f8bf16((A, As), (B, Bs), C)
 
     def deep_gemm_fp8_fp8_bf16_nt_fake(
         A: torch.Tensor,
@@ -280,6 +276,7 @@ def sglang_per_token_group_quant_fp8(
     eps: float = 1e-10,
     column_major_scales: bool = False,
     scale_tma_aligned: bool = False,
+    scale_ue8m0: bool = False,
 ):
     assert (
         x.shape[-1] % group_size == 0
@@ -287,8 +284,21 @@ def sglang_per_token_group_quant_fp8(
     assert x.is_contiguous(), "`x` is not contiguous"
 
     x_q = torch.empty_like(x, device=x.device, dtype=fp8_dtype)
-    if column_major_scales:
+    if scale_ue8m0:
+        assert column_major_scales and scale_tma_aligned
+        x_q_mn, x_q_k = x.shape
+        x_s_mn, x_s_k = x_q_mn, x_q_k // 128
+        aligned_mn = align(x_s_mn, 4)
+        aligned_k = align(x_s_k, 4)
+        # TODO(FIXME): Fix cuda kernel and recover here to empty.
+        x_s = torch.zeros(
+            (aligned_k // 4, aligned_mn),
+            device=x.device,
+            dtype=torch.int,
+        ).transpose(0, 1)[:x_s_mn, :]
+    elif column_major_scales:
         if scale_tma_aligned:
+            # TODO extract "align" function
             # aligned to 4 * sizeof(float)
             aligned_size = (x.shape[-2] + 3) // 4 * 4
             x_s = torch.empty(
@@ -309,7 +319,9 @@ def sglang_per_token_group_quant_fp8(
             dtype=torch.float32,
         )
     if x.shape[0] > 0:
-        sgl_per_token_group_quant_fp8(x, x_q, x_s, group_size, eps, fp8_min, fp8_max)
+        sgl_per_token_group_quant_fp8(
+            x, x_q, x_s, group_size, eps, fp8_min, fp8_max, scale_ue8m0
+        )
 
     return x_q, x_s
 
@@ -754,7 +766,15 @@ def prepare_block_fp8_matmul_inputs(
     assert A.shape[-1] == B.shape[-1]
     assert A.shape[:-1] == As.shape[:-1]
     assert A.is_contiguous()
-    assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+
+    if As.dtype == torch.float:
+        assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    elif As.dtype == torch.int:
+        assert (
+            triton.cdiv(triton.cdiv(A.shape[-1], block_k), 4) == As.shape[-1]
+        ), f"{A.shape=} {As.shape=} {block_size=}"
+    else:
+        raise NotImplementedError
 
     M = A.numel() // A.shape[-1]
 
@@ -762,8 +782,17 @@ def prepare_block_fp8_matmul_inputs(
     assert B.is_contiguous()
     assert Bs.ndim == 2
     N, K = B.shape
-    assert triton.cdiv(N, block_n) == Bs.shape[0]
-    assert triton.cdiv(K, block_k) == Bs.shape[1]
+
+    if Bs.dtype == torch.float:
+        assert triton.cdiv(N, block_n) == Bs.shape[0]
+        assert triton.cdiv(K, block_k) == Bs.shape[1]
+    elif Bs.dtype == torch.int:
+        assert N == Bs.shape[0], f"{B.shape=} {Bs.shape=} {block_size=}"
+        assert (
+            triton.cdiv(triton.cdiv(K, block_k), 4) == Bs.shape[1]
+        ), f"{B.shape=} {Bs.shape=} {block_size=}"
+    else:
+        raise NotImplementedError
 
     C_shape = A.shape[:-1] + (N,)
     C = A.new_empty(C_shape, dtype=output_dtype)
@@ -782,12 +811,12 @@ def w8a8_block_fp8_matmul_deepgemm(
     M, N, K, C = prepare_block_fp8_matmul_inputs(A, B, As, Bs, block_size, output_dtype)
 
     # Deepgemm only supports output tensor type as bfloat16
-    assert C.dtype == torch.bfloat16 and _ENABLE_JIT_DEEPGEMM
+    assert C.dtype == torch.bfloat16 and deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
 
     if supports_custom_op():
         torch.ops.sglang.deep_gemm_fp8_fp8_bf16_nt(A, As, B, Bs, C)
     else:
-        deep_gemm_gemm_nt_f8f8bf16((A, As), (B, Bs), C)
+        deep_gemm_wrapper.gemm_nt_f8f8bf16((A, As), (B, Bs), C)
 
     return C
 
@@ -881,7 +910,7 @@ def w8a8_block_fp8_matmul(
     block_size: List[int],
     output_dtype: torch.dtype = torch.float16,
 ) -> torch.Tensor:
-    if output_dtype == torch.bfloat16 and _ENABLE_JIT_DEEPGEMM:
+    if output_dtype == torch.bfloat16 and deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
         return w8a8_block_fp8_matmul_deepgemm(
             A, B, As, Bs, block_size, output_dtype=output_dtype
         )

sglang/srt/layers/quantization/fp8_utils.py

@@ -1,9 +1,10 @@
-import os
-from curses import flash
 from typing import Callable, List, Optional, Tuple
 
+import einops
 import torch
 
+from sglang.math_utils import align
+from sglang.srt.layers.quantization import deep_gemm_wrapper
 from sglang.srt.layers.quantization.fp8_kernel import sglang_per_token_group_quant_fp8
 from sglang.srt.layers.utils import is_sm100_supported
 
@@ -14,7 +15,6 @@ try:
 except ImportError:
     VLLM_AVAILABLE = False
 
-from sglang.srt.layers.quantization.deep_gemm import _ENABLE_JIT_DEEPGEMM
 from sglang.srt.layers.quantization.fp8_kernel import (
     fp8_dtype,
     fp8_max,
@@ -137,7 +137,7 @@ def dispatch_w8a8_block_fp8_linear() -> Callable:
         return cutlass_w8a8_block_fp8_linear_with_fallback
     elif _use_aiter:
         return aiter_w8a8_block_fp8_linear
-    elif _ENABLE_JIT_DEEPGEMM:
+    elif deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
         return deepgemm_w8a8_block_fp8_linear_with_fallback
     else:
         return triton_w8a8_block_fp8_linear
@@ -238,7 +238,14 @@ def deepgemm_w8a8_block_fp8_linear_with_fallback(
         block_size[1],
         column_major_scales=True,
         scale_tma_aligned=True,
+        scale_ue8m0=deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0,
     )
+
+    # NOTE(alcanderian): Useless when scale is packed to int32
+    # if get_bool_env_var("SGLANG_W8A8_DEEPGEMM_SANITY_CHECK_UE8M0"):
+    #     _check_ue8m0("x_scale", x_scale)
+    #     _check_ue8m0("weight_scale", ws)
+
     output = w8a8_block_fp8_matmul_deepgemm(
         q_input, weight, x_scale, weight_scale, block_size, output_dtype=output_dtype
     )
@@ -247,6 +254,11 @@ def deepgemm_w8a8_block_fp8_linear_with_fallback(
     return output.to(dtype=output_dtype).view(*output_shape)
 
 
+def _check_ue8m0(name, x):
+    x_ceil = ceil_to_ue8m0(x)
+    assert torch.all(x == x_ceil), f"{name=} {x=} {x_ceil=}"
+
+
 def aiter_w8a8_block_fp8_linear(
     input: torch.Tensor,
     weight: torch.Tensor,
@@ -369,27 +381,80 @@ def block_quant_dequant(
     The output is an unquantized tensor with dtype.
     """
     block_n, block_k = block_size[0], block_size[1]
-    n, k = x_q_block.shape
-    n_tiles = (n + block_n - 1) // block_n
-    k_tiles = (k + block_k - 1) // block_k
-    assert n_tiles == x_s.shape[0]
-    assert k_tiles == x_s.shape[1]
+    *_, n, k = x_q_block.shape
 
-    x_dq_block = torch.empty_like(x_q_block, dtype=dtype)
+    # ... n_scale k_scale -> ... (n_scale block_n) (k_scale block_k)
+    x_scale_repeat = x_s.repeat_interleave(block_n, dim=-2).repeat_interleave(
+        block_k, dim=-1
+    )
+    x_scale_repeat = x_scale_repeat[..., :n, :k]
+
+    return (x_q_block.to(torch.float32) * x_scale_repeat).to(dtype)
+
+
+def requant_weight_ue8m0_inplace(weight, weight_scale_inv, weight_block_size):
+    assert isinstance(weight, torch.nn.Parameter)
+    assert isinstance(weight_scale_inv, torch.nn.Parameter)
+    weight.data, weight_scale_inv.data = _requant_weight_ue8m0(
+        weight, weight_scale_inv, weight_block_size
+    )
+
+
+def _requant_weight_ue8m0(
+    weight: torch.Tensor,
+    weight_scale_inv: torch.Tensor,
+    weight_block_size: List[int],
+):
+    assert weight_block_size == [128, 128]
+
+    *_, n, k = weight.shape
+
+    weight_dequant = block_quant_dequant(
+        weight,
+        weight_scale_inv,
+        weight_block_size,
+        torch.bfloat16,
+    )
+
+    weight_dequant_flat = weight_dequant.view((-1, k))
+    out_w_flat, out_s_flat = per_block_cast_to_fp8(weight_dequant_flat)
+
+    out_w = out_w_flat.view(weight.shape)
+    out_s = out_s_flat.view(weight_scale_inv.shape)
+
+    # NOTE copy and modified from DeepGEMM
+    def _transform_scale(sf, mn: int):
+        import deep_gemm.utils.layout
+
+        sf = sf.index_select(-2, torch.arange(mn, device=sf.device) // 128)
+        sf = deep_gemm.utils.layout.get_col_major_tma_aligned_packed_tensor(sf)
+        return sf
+
+    out_s = _transform_scale(out_s, mn=out_w.shape[-2])
+
+    return out_w, out_s
+
+
+# COPIED FROM DeepGEMM
+def per_block_cast_to_fp8(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    assert x.dim() == 2
+    m, n = x.shape
+    x_padded = torch.zeros(
+        (align(m, 128), align(n, 128)), dtype=x.dtype, device=x.device
+    )
+    x_padded[:m, :n] = x
+    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
+    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
+    sf = ceil_to_ue8m0(x_amax / 448.0)
+    x_scaled = (x_view * (1.0 / sf)).to(torch.float8_e4m3fn)
+    return x_scaled.view_as(x_padded)[:m, :n].contiguous(), sf.view(
+        x_view.size(0), x_view.size(2)
+    )
 
-    for j in range(n_tiles):
-        for i in range(k_tiles):
-            x_q_block_tile = x_q_block[
-                j * block_n : min((j + 1) * block_n, n),
-                i * block_k : min((i + 1) * block_k, k),
-            ]
-            x_dq_block_tile = x_dq_block[
-                j * block_n : min((j + 1) * block_n, n),
-                i * block_k : min((i + 1) * block_k, k),
-            ]
-            x_dq_block_tile[:, :] = x_q_block_tile.to(torch.float32) * x_s[j][i]
 
-    return x_dq_block
+# COPIED FROM DeepGEMM
+def ceil_to_ue8m0(x: torch.Tensor):
+    return torch.pow(2.0, torch.ceil(torch.log2(x.abs())))
 
 
 def channel_quant_to_tensor_quant(