sglang 0.4.0.post2__py3-none-any.whl → 0.4.1.post1__py3-none-any.whl

sglang/srt/layers/quantization/fp8_kernel.py

@@ -0,0 +1,292 @@
+# Copyright 2024 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+from typing import List, Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _per_token_group_quant_fp8(
+    # Pointers to inputs and output
+    y_ptr,
+    y_q_ptr,
+    y_s_ptr,
+    # Stride of input
+    y_stride,
+    # Collums of input
+    N,
+    # Avoid to divide zero
+    eps,
+    # Information for float8
+    fp8_min,
+    fp8_max,
+    # Meta-parameters
+    BLOCK: tl.constexpr,
+):
+    """A Triton-accelerated function to perform per-token-group quantization on a
+    tensor.
+
+    This function converts the tensor values into float8 values.
+    """
+    # Map the program id to the row of X and Y it should compute.
+    g_id = tl.program_id(0)
+    y_ptr += g_id * y_stride
+    y_q_ptr += g_id * y_stride
+    y_s_ptr += g_id
+
+    cols = tl.arange(0, BLOCK)  # N <= BLOCK
+    mask = cols < N
+
+    y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
+    # Quant
+    _absmax = tl.maximum(tl.max(tl.abs(y)), eps)
+    y_s = _absmax / fp8_max
+    y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
+
+    tl.store(y_q_ptr + cols, y_q, mask=mask)
+    tl.store(y_s_ptr, y_s)
+
+
+def per_token_group_quant_fp8(
+    x: torch.Tensor,
+    group_size: int,
+    eps: float = 1e-10,
+    dtype: torch.dtype = torch.float8_e4m3fn,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Function to perform per-token-group quantization on an input tensor `x`.
+
+    It converts the tensor values into signed float8 values and returns the
+    quantized tensor along with the scaling factor used for quantization.
+
+    Args:
+        x: The input tenosr with ndim >= 2.
+        group_size: The group size used for quantization.
+        eps: The minimum to avoid dividing zero.
+        dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn` is supported for now.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the scaling factor for quantization.
+    """
+    assert (
+        x.shape[-1] % group_size == 0
+    ), "the last dimension of `x` cannot be divisible by `group_size`"
+    assert x.is_contiguous(), "`x` is not contiguous"
+
+    finfo = torch.finfo(dtype)
+    fp8_min = finfo.min
+    fp8_max = finfo.max
+
+    x_q = torch.empty_like(x, device=x.device, dtype=dtype)
+    M = x.numel() // group_size
+    N = group_size
+    x_s = torch.empty(
+        x.shape[:-1] + (x.shape[-1] // group_size,),
+        device=x.device,
+        dtype=torch.float32,
+    )
+
+    BLOCK = triton.next_power_of_2(N)
+    # heuristics for number of warps
+    num_warps = min(max(BLOCK // 256, 1), 8)
+    num_stages = 1
+    _per_token_group_quant_fp8[(M,)](
+        x,
+        x_q,
+        x_s,
+        group_size,
+        N,
+        eps,
+        fp8_min=fp8_min,
+        fp8_max=fp8_max,
+        BLOCK=BLOCK,
+        num_warps=num_warps,
+        num_stages=num_stages,
+    )
+
+    return x_q, x_s
+
+
+@triton.jit
+def _w8a8_block_fp8_matmul(
+    # Pointers to inputs and output
+    A,
+    B,
+    C,
+    As,
+    Bs,
+    # Shape for matmul
+    M,
+    N,
+    K,
+    # Block size for block-wise quantization
+    group_n,
+    group_k,
+    # Stride for inputs and output
+    stride_am,
+    stride_ak,
+    stride_bk,
+    stride_bn,
+    stride_cm,
+    stride_cn,
+    stride_As_m,
+    stride_As_k,
+    stride_Bs_k,
+    stride_Bs_n,
+    # Meta-parameters
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,
+    GROUP_SIZE_M: tl.constexpr,
+):
+    """Triton-accelerated function used to perform linear operations (dot
+    product) on input tensors `A` and `B` with block-wise quantization, and store the result in output
+    tensor `C`.
+    """
+
+    pid = tl.program_id(axis=0)
+    num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+    num_pid_in_group = GROUP_SIZE_M * num_pid_n
+    group_id = pid // num_pid_in_group
+    first_pid_m = group_id * GROUP_SIZE_M
+    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
+    pid_m = first_pid_m + (pid % group_size_m)
+    pid_n = (pid % num_pid_in_group) // group_size_m
+
+    offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
+    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
+    b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
+
+    As_ptrs = As + offs_am * stride_As_m
+    offs_bsn = offs_bn // group_n
+    Bs_ptrs = Bs + offs_bsn * stride_Bs_n
+
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+        a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
+        b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
+
+        k_start = k * BLOCK_SIZE_K
+        offs_ks = k_start // group_k
+        a_s = tl.load(As_ptrs + offs_ks * stride_As_k)
+        b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k)
+
+        accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        b_ptrs += BLOCK_SIZE_K * stride_bk
+
+    if C.dtype.element_ty == tl.bfloat16:
+        c = accumulator.to(tl.bfloat16)
+    elif C.dtype.element_ty == tl.float16:
+        c = accumulator.to(tl.float16)
+    else:
+        c = accumulator.to(tl.float32)
+
+    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
+    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, c, mask=c_mask)
+
+
+def w8a8_block_fp8_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    As: torch.Tensor,
+    Bs: torch.Tensor,
+    block_size: List[int],
+    output_dtype: torch.dtype = torch.float16,
+) -> torch.Tensor:
+    """This function performs matrix multiplication with block-wise quantization.
+
+    It takes two input tensors `A` and `B` with scales `As` and `Bs`.
+    The output is returned in the specified `output_dtype`.
+
+    Args:
+        A: The input tensor, e.g., activation.
+        B: The input tensor, e.g., weight.
+        As: The per-token-group quantization scale for `A`.
+        Bs: The per-block quantization scale for `B`.
+        block_size: The block size for per-block quantization. It should be 2-dim, e.g., [128, 128].
+        output_dytpe: The dtype of the returned tensor.
+
+    Returns:
+        torch.Tensor: The result of matmul.
+    """
+    assert len(block_size) == 2
+    block_n, block_k = block_size[0], block_size[1]
+
+    assert A.shape[-1] == B.shape[-1]
+    assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
+    assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    M = A.numel() // A.shape[-1]
+
+    assert B.ndim == 2 and B.is_contiguous() and 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]
+
+    C_shape = A.shape[:-1] + (N,)
+    C = A.new_empty(C_shape, dtype=output_dtype)
+
+    # TODO(HandH1998):
+    # BLOCK_SIZE_M, BLOCK_SIZE_K, BLOCK_SIZE_N can be optimized.
+    # BLOCK_SIZE_K must be divisable by block_k
+    # BLOCK_SIZE_N and BLOCK_SIZE_M has no requirements
+    BLOCK_SIZE_M = 128
+    if M < BLOCK_SIZE_M:
+        BLOCK_SIZE_M = triton.next_power_of_2(M)
+        BLOCK_SIZE_M = max(BLOCK_SIZE_M, 16)
+    BLOCK_SIZE_K = block_k
+    assert block_k % BLOCK_SIZE_K == 0
+    BLOCK_SIZE_N = block_n
+
+    def grid(META):
+        return (
+            triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
+        )
+
+    _w8a8_block_fp8_matmul[grid](
+        A,
+        B,
+        C,
+        As,
+        Bs,
+        M,
+        N,
+        K,
+        block_n,
+        block_k,
+        A.stride(-2),
+        A.stride(-1),
+        B.stride(1),
+        B.stride(0),
+        C.stride(-2),
+        C.stride(-1),
+        As.stride(-2),
+        As.stride(-1),
+        Bs.stride(1),
+        Bs.stride(0),
+        BLOCK_SIZE_M=BLOCK_SIZE_M,
+        BLOCK_SIZE_N=BLOCK_SIZE_N,
+        BLOCK_SIZE_K=BLOCK_SIZE_K,
+        GROUP_SIZE_M=8,
+    )
+
+    return C

sglang/srt/layers/quantization/fp8_utils.py

@@ -1,6 +1,12 @@
-from typing import Optional, Tuple
+from typing import List, Optional, Tuple
 
 import torch
+from vllm.model_executor.parameter import RowvLLMParameter, _ColumnvLLMParameter
+
+from sglang.srt.layers.quantization.fp8_kernel import (
+    per_token_group_quant_fp8,
+    w8a8_block_fp8_matmul,
+)
 
 
 def normalize_e4m3fn_to_e4m3fnuz(
@@ -25,3 +31,86 @@ def normalize_e4m3fn_to_e4m3fnuz(
     if input_scale is not None:
         input_scale = input_scale * 2.0
     return weight, weight_scale, input_scale
+
+
+def apply_w8a8_block_fp8_linear(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    block_size: List[int],
+    weight_scale: torch.Tensor,
+    input_scale: Optional[torch.Tensor] = None,
+    bias: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    assert input_scale is None
+    # View input as 2D matrix for fp8 methods
+    input_2d = input.view(-1, input.shape[-1])
+    output_shape = [*input.shape[:-1], weight.shape[0]]
+
+    q_input, x_scale = per_token_group_quant_fp8(input_2d, block_size[1])
+    output = w8a8_block_fp8_matmul(
+        q_input, weight, x_scale, weight_scale, block_size, output_dtype=input.dtype
+    )
+
+    if bias is not None:
+        output = output + bias
+    return output.to(dtype=input.dtype).view(*output_shape)
+
+
+def input_to_float8(
+    x: torch.Tensor, dtype: torch.dtype = torch.float8_e4m3fn
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """This function quantizes input values to float8 values with tensor-wise quantization."""
+    finfo = torch.finfo(dtype)
+    min_val, max_val = x.aminmax()
+    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    scale = finfo.max / amax
+    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
+    return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
+
+
+def block_quant_to_tensor_quant(
+    x_q_block: torch.Tensor,
+    x_s: torch.Tensor,
+    block_size: List[int],
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """This function converts block-wise quantization to tensor-wise quantization.
+    The inputs are block-wise quantization tensor `x_q_block`, block-wise quantization scale
+    and the block size.
+    The outputs are tensor-wise quantization tensor and tensor-wise quantization scale.
+    Note only float8 is supported for now.
+    """
+    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]
+
+    x_dq_block = x_q_block.to(torch.float32)
+
+    x_dq_block_tiles = [
+        [
+            x_dq_block[
+                j * block_n : min((j + 1) * block_n, n),
+                i * block_k : min((i + 1) * block_k, k),
+            ]
+            for i in range(k_tiles)
+        ]
+        for j in range(n_tiles)
+    ]
+
+    for i in range(k_tiles):
+        for j in range(n_tiles):
+            x_dq_block_tiles[j][i][:, :] = x_dq_block_tiles[j][i] * x_s[j][i]
+
+    x_q_tensor, scale = input_to_float8(x_dq_block, dtype=x_q_block.dtype)
+    return x_q_tensor, scale
+
+
+class BlockQuantScaleParameter(_ColumnvLLMParameter, RowvLLMParameter):
+    """
+    Parameter class for weight scales loaded for weights with
+    block-wise quantization. Uses both column and row parallelism.
+    """
+
+    pass

sglang/srt/layers/torchao_utils.py

@@ -2,8 +2,14 @@
 Common utilities for torchao.
 """
 
+import logging
+import os
+import pwd
+
 import torch
 
+logger = logging.getLogger(__name__)
+
 
 def apply_torchao_config_to_model(
     model: torch.nn.Module, torchao_config: str, filter_fn=None
@@ -50,27 +56,17 @@ def apply_torchao_config_to_model(
     elif "gemlite" in torchao_config:
         # gemlite-<packing_bitwidth>-<bit_width>-<group_size> or
         # gemlite-<bit_width>-<group_size> (packing_bitwidth defaults to 32)
-        import os
-        import pwd
-
-        import gemlite
-        from gemlite.core import GemLiteLinearTriton, set_autotune
-
-        try:
-            from torchao.quantization import gemlite_uintx_weight_only
-        except:
-            print(
-                f"import `gemlite_uintx_weight_only` failed, please use torchao nightly to use gemlite quantization"
-            )
-            return model
+        from gemlite.core import GemLiteLinearTriton
+        from torchao.quantization import gemlite_uintx_weight_only
 
         _quant_args = torchao_config.split("-")
         bit_width = int(_quant_args[-2])
         group_size = None if _quant_args[-1] == "None" else int(_quant_args[-1])
+
         try:
             packing_bitwidth = int(_quant_args[-3])
-        except:
-            # if only 2 inputs found, use default value
+        except (ValueError, IndexError):
+            # if only 2 inputs found or conversion fails, use default value
             packing_bitwidth = 32
 
         quantize_(

sglang/srt/managers/schedule_batch.py

@@ -479,8 +479,22 @@ class Req:
 
         return True
 
+    def reset_for_retract(self):
+        self.prefix_indices = []
+        self.last_node = None
+        self.extend_input_len = 0
+        self.is_retracted = True
+
+        # For incremental logprobs
+        # TODO: Fix the `logprob_start_len`
+        self.last_update_decode_tokens = 0
+        self.logprob_start_len = 10**9
+
     def __repr__(self):
-        return f"rid(n={self.rid}, " f"input_ids={self.origin_input_ids}, "
+        return (
+            f"rid(n={self.rid}, "
+            f"input_ids={self.origin_input_ids}, output_ids={self.output_ids}"
+        )
 
 
 bid = 0
@@ -894,15 +908,7 @@ class ScheduleBatch:
                 )
                 residual_size = max(0, residual_size)
                 self.tree_cache.evict(residual_size, self.token_to_kv_pool.free)
-
-            req.prefix_indices = []
-            req.last_node = None
-            req.extend_input_len = 0
-            req.is_retracted = True
-
-            # For incremental logprobs
-            req.last_update_decode_tokens = 0
-            req.logprob_start_len = 10**9
+            req.reset_for_retract()
 
         self.filter_batch(keep_indices=sorted_indices)
 

sglang/srt/managers/schedule_policy.py

@@ -248,7 +248,7 @@ class PrefillAdder:
         self.can_run_list.append(req)
 
         self._prefill_one_req(
-            len(req.prefix_indices),
+            0,
             req.extend_input_len,
             (
                 min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS_ESTIMATION)

sglang/srt/managers/scheduler.py

@@ -22,7 +22,7 @@ import warnings
 from collections import deque
 from concurrent import futures
 from types import SimpleNamespace
-from typing import List, Optional
+from typing import Callable, Dict, List, Optional, Tuple
 
 import psutil
 import setproctitle
@@ -260,7 +260,7 @@ class Scheduler:
         self.current_stream = torch.get_device_module(self.device).current_stream()
 
         # Session info
-        self.sessions = {}
+        self.sessions: Dict[str, Session] = {}
 
         # Init chunked prefill
         self.chunked_prefill_size = server_args.chunked_prefill_size
@@ -468,9 +468,6 @@ class Scheduler:
                 self.send_to_tokenizer.send_pyobj(
                     UpdateWeightFromDiskReqOutput(success, message)
                 )
-            elif isinstance(recv_req, GetWeightsByNameReqInput):
-                parameter = self.get_weights_by_name(recv_req)
-                self.send_to_tokenizer.send_pyobj(GetWeightsByNameReqOutput(parameter))
             elif isinstance(recv_req, InitWeightsUpdateGroupReqInput):
                 success, message = self.init_weights_update_group(recv_req)
                 self.send_to_tokenizer.send_pyobj(
@@ -565,7 +562,7 @@ class Scheduler:
 
         if req.logprob_start_len == -1:
             # By default, only return the logprobs for output tokens
-            req.logprob_start_len = len(recv_req.input_ids) - 1
+            req.logprob_start_len = len(req.origin_input_ids) - 1
 
         # Truncate prompts that are too long
         if len(req.origin_input_ids) > self.max_req_input_len:
@@ -589,12 +586,15 @@ class Scheduler:
         if (
             req.sampling_params.json_schema is not None
             or req.sampling_params.regex is not None
+            or req.sampling_params.ebnf is not None
         ):
             assert self.grammar_backend is not None
             if req.sampling_params.json_schema is not None:
                 key = ("json", req.sampling_params.json_schema)
             elif req.sampling_params.regex is not None:
                 key = ("regex", req.sampling_params.regex)
+            elif req.sampling_params.ebnf is not None:
+                key = ("ebnf", req.sampling_params.ebnf)
 
             req.grammar = self.grammar_backend.get_cached_value(key)
             if not req.grammar:
@@ -629,16 +629,13 @@ class Scheduler:
         self.waiting_queue.append(req)
 
     def log_prefill_stats(self, adder, can_run_list, running_bs, has_being_chunked):
-        if isinstance(self.tree_cache, RadixCache):
-            self.tree_cache_metrics["total"] += (
-                adder.log_input_tokens + adder.log_hit_tokens
-            ) / 10**9
-            self.tree_cache_metrics["hit"] += (adder.log_hit_tokens) / 10**9
-            tree_cache_hit_rate = (
-                self.tree_cache_metrics["hit"] / self.tree_cache_metrics["total"]
-            )
-        else:
-            tree_cache_hit_rate = 0.0
+        self.tree_cache_metrics["total"] += (
+            adder.log_input_tokens + adder.log_hit_tokens
+        ) / 10**9
+        self.tree_cache_metrics["hit"] += (adder.log_hit_tokens) / 10**9
+        tree_cache_hit_rate = (
+            self.tree_cache_metrics["hit"] / self.tree_cache_metrics["total"]
+        )
 
         num_used = self.max_total_num_tokens - (
             self.token_to_kv_pool.available_size() + self.tree_cache.evictable_size()