sglang 0.4.1.post3__py3-none-any.whl → 0.4.1.post5__py3-none-any.whl

sglang/srt/layers/logits_processor.py

@@ -17,6 +17,8 @@ import dataclasses
 from typing import List, Optional, Union
 
 import torch
+import triton
+import triton.language as tl
 from torch import nn
 from vllm.distributed import (
     get_tensor_model_parallel_world_size,
@@ -33,76 +35,72 @@ from sglang.srt.model_executor.forward_batch_info import (
 
 @dataclasses.dataclass
 class LogitsProcessorOutput:
+    ## Part 1: This part will be assigned in python/sglang/srt/layers/logits_processor.py::LogitsProcessor
     # The logits of the next tokens.       shape: [#seq, vocab_size]
     next_token_logits: torch.Tensor
-    # The logprobs of the next tokens.     shape: [#seq, vocab_size]
-    next_token_logprobs: torch.Tensor = None
+    # Used by speculative decoding (EAGLE)
+    # The last hidden layers
+    hidden_states: Optional[torch.Tensor] = None
 
+    ## Part 2: This part will be assigned in python/sglang/srt/layers/sampler.py::Sampler
+    # The logprobs of the next tokens.                              shape: [#seq]
+    next_token_logprobs: Optional[torch.Tensor] = None
+    # The logprobs and ids of the top-k tokens in output positions. shape: [#seq, k]
+    next_token_top_logprobs_val: Optional[List] = None
+    next_token_top_logprobs_idx: Optional[List] = None
+
+    ## Part 3: Prefill-only. This part will be assigned in python/sglang/srt/layers/logits_processor.py::LogitsProcessor
     # The normlaized logprobs of prompts.  shape: [#seq]
     normalized_prompt_logprobs: torch.Tensor = None
-    # The logprobs of input tokens.        shape: [#token, vocab_size]
+    # The logprobs of input tokens.        shape: [#token]
     input_token_logprobs: torch.Tensor = None
-
-    # The logprob and id of the top-k tokens in input positions.  shape [#seq, #token, k]
+    # The logprobs and ids of the top-k tokens in input positions.  shape: [#seq, #token, k]
     input_top_logprobs_val: List = None
     input_top_logprobs_idx: List = None
-    # The logprob and id of the top-k tokens in output positions. shape [#seq, #token, k]
-    output_top_logprobs_val: List = None
-    output_top_logprobs_idx: List = None
-
-    # Used by speculative decoding (EAGLE)
-    # The output of transformer layers
-    hidden_states: Optional[torch.Tensor] = None
 
 
 @dataclasses.dataclass
 class LogitsMetadata:
     forward_mode: ForwardMode
-    top_logprobs_nums: Optional[List[int]]
-
-    return_logprob: bool = False
-    return_top_logprob: bool = False
+    capture_hidden_mode: CaptureHiddenMode = CaptureHiddenMode.NULL
 
+    extend_return_logprob: bool = False
+    extend_return_top_logprob: bool = False
     extend_seq_lens: Optional[torch.Tensor] = None
     extend_seq_lens_cpu: Optional[List[int]] = None
-
     extend_logprob_start_lens_cpu: Optional[List[int]] = None
     extend_logprob_pruned_lens_cpu: Optional[List[int]] = None
-
-    capture_hidden_mode: CaptureHiddenMode = CaptureHiddenMode.NULL
+    top_logprobs_nums: Optional[List[int]] = None
 
     @classmethod
     def from_forward_batch(cls, forward_batch: ForwardBatch):
-        extend_logprob_pruned_lens_cpu = None
-
-        if forward_batch.return_logprob:
-            return_top_logprob = any(x > 0 for x in forward_batch.top_logprobs_nums)
-            if forward_batch.forward_mode.is_extend():
-                extend_logprob_pruned_lens_cpu = [
-                    extend_len - start_len
-                    for extend_len, start_len in zip(
-                        forward_batch.extend_seq_lens_cpu,
-                        forward_batch.extend_logprob_start_lens_cpu,
-                    )
-                ]
-        else:
-            return_top_logprob = False
-
-        if forward_batch.spec_info:
-            capture_hidden_mode = forward_batch.spec_info.capture_hidden_mode
+        if forward_batch.forward_mode.is_extend() and forward_batch.return_logprob:
+            extend_return_logprob = True
+            extend_return_top_logprob = any(
+                x > 0 for x in forward_batch.top_logprobs_nums
+            )
+            extend_logprob_pruned_lens_cpu = [
+                extend_len - start_len
+                for extend_len, start_len in zip(
+                    forward_batch.extend_seq_lens_cpu,
+                    forward_batch.extend_logprob_start_lens_cpu,
+                )
+            ]
         else:
-            capture_hidden_mode = CaptureHiddenMode.NULL
+            extend_return_logprob = extend_return_top_logprob = (
+                extend_logprob_pruned_lens_cpu
+            ) = False
 
         return cls(
             forward_mode=forward_batch.forward_mode,
-            top_logprobs_nums=forward_batch.top_logprobs_nums,
-            return_logprob=forward_batch.return_logprob,
-            return_top_logprob=return_top_logprob,
+            capture_hidden_mode=forward_batch.capture_hidden_mode,
+            extend_return_logprob=extend_return_logprob,
+            extend_return_top_logprob=extend_return_top_logprob,
             extend_seq_lens=forward_batch.extend_seq_lens,
             extend_seq_lens_cpu=forward_batch.extend_seq_lens_cpu,
             extend_logprob_start_lens_cpu=forward_batch.extend_logprob_start_lens_cpu,
             extend_logprob_pruned_lens_cpu=extend_logprob_pruned_lens_cpu,
-            capture_hidden_mode=capture_hidden_mode,
+            top_logprobs_nums=forward_batch.top_logprobs_nums,
         )
 
 
@@ -119,6 +117,11 @@ class LogitsProcessor(nn.Module):
         self.final_logit_softcapping = getattr(
             self.config, "final_logit_softcapping", None
         )
+        if (
+            self.final_logit_softcapping is not None
+            and self.final_logit_softcapping < 0
+        ):
+            self.final_logit_softcapping = None
 
     def forward(
         self,
@@ -129,7 +132,6 @@ class LogitsProcessor(nn.Module):
     ):
         if isinstance(logits_metadata, ForwardBatch):
             logits_metadata = LogitsMetadata.from_forward_batch(logits_metadata)
-        assert isinstance(logits_metadata, LogitsMetadata)
 
         # Get the last hidden states and last logits for the next token prediction
         if (
@@ -142,18 +144,13 @@ class LogitsProcessor(nn.Module):
             last_index = torch.cumsum(logits_metadata.extend_seq_lens, dim=0) - 1
             last_hidden = hidden_states[last_index]
 
+        # Compute logits
         last_logits = self._get_logits(last_hidden, lm_head)
-        if self.do_tensor_parallel_all_gather:
-            last_logits = tensor_model_parallel_all_gather(last_logits)
-        last_logits = last_logits[:, : self.config.vocab_size].float()
-
-        if self.final_logit_softcapping:
-            last_logits.div_(self.final_logit_softcapping)
-            torch.tanh(last_logits, out=last_logits)
-            last_logits.mul_(self.final_logit_softcapping)
-
-        # Return only last_logits if logprob is not requested
-        if not logits_metadata.return_logprob:
+        if (
+            not logits_metadata.extend_return_logprob
+            or logits_metadata.capture_hidden_mode.need_capture()
+        ):
+            # Decode mode or extend mode without return_logprob.
             return LogitsProcessorOutput(
                 next_token_logits=last_logits,
                 hidden_states=(
@@ -167,95 +164,60 @@ class LogitsProcessor(nn.Module):
                 ),
             )
         else:
-            last_logprobs = self.compute_temp_top_p_normalized_logprobs(
-                last_logits, logits_metadata
+            # Slice the requested tokens to compute logprob
+            pt, pruned_states, pruned_input_ids = 0, [], []
+            for start_len, extend_len in zip(
+                logits_metadata.extend_logprob_start_lens_cpu,
+                logits_metadata.extend_seq_lens_cpu,
+            ):
+                pruned_states.append(hidden_states[pt + start_len : pt + extend_len])
+                pruned_input_ids.append(input_ids[pt + start_len : pt + extend_len])
+                pt += extend_len
+
+            # Compute the logits of all required tokens
+            pruned_states = torch.cat(pruned_states)
+            del hidden_states
+            input_token_logits = self._get_logits(pruned_states, lm_head)
+            del pruned_states
+
+            # Normalize the logprob w/o temperature, top-p
+            input_logprobs = input_token_logits
+            input_logprobs = self.compute_temp_top_p_normalized_logprobs(
+                input_logprobs, logits_metadata
             )
 
-            if logits_metadata.forward_mode.is_decode():
-                if logits_metadata.return_top_logprob:
-                    output_top_logprobs_val, output_top_logprobs_idx = (
-                        self.get_top_logprobs(last_logprobs, logits_metadata)[2:4]
-                    )
-                else:
-                    output_top_logprobs_val = output_top_logprobs_idx = None
-                return LogitsProcessorOutput(
-                    next_token_logits=last_logits,
-                    next_token_logprobs=last_logprobs,
-                    output_top_logprobs_val=output_top_logprobs_val,
-                    output_top_logprobs_idx=output_top_logprobs_idx,
-                )
+            # Get the logprob of top-k tokens
+            if logits_metadata.extend_return_top_logprob:
+                (
+                    input_top_logprobs_val,
+                    input_top_logprobs_idx,
+                ) = self.get_top_logprobs(input_logprobs, logits_metadata)
             else:
-                # Slice the requested tokens to compute logprob
-                pt, states, pruned_input_ids = 0, [], []
-                for start_len, extend_len in zip(
-                    logits_metadata.extend_logprob_start_lens_cpu,
-                    logits_metadata.extend_seq_lens_cpu,
-                ):
-                    states.append(hidden_states[pt + start_len : pt + extend_len])
-                    pruned_input_ids.append(input_ids[pt + start_len : pt + extend_len])
-                    pt += extend_len
-
-                # Compute the logits and logprobs for all required tokens
-                states = torch.cat(states, dim=0)
-                all_logits = self._get_logits(states, lm_head)
-                if self.do_tensor_parallel_all_gather:
-                    all_logits = tensor_model_parallel_all_gather(all_logits)
-
-                # The LM head's weights may be zero-padded for parallelism. Remove any
-                # extra logits that this padding may have produced.
-                all_logits = all_logits[:, : self.config.vocab_size].float()
-
-                if self.final_logit_softcapping:
-                    all_logits.div_(self.final_logit_softcapping)
-                    torch.tanh(all_logits, out=all_logits)
-                    all_logits.mul_(self.final_logit_softcapping)
-
-                all_logprobs = all_logits
-                del all_logits, hidden_states
-
-                all_logprobs = self.compute_temp_top_p_normalized_logprobs(
-                    all_logprobs, logits_metadata
-                )
-
-                # Get the logprob of top-k tokens
-                if logits_metadata.return_top_logprob:
-                    (
-                        input_top_logprobs_val,
-                        input_top_logprobs_idx,
-                        output_top_logprobs_val,
-                        output_top_logprobs_idx,
-                    ) = self.get_top_logprobs(all_logprobs, logits_metadata)
-                else:
-                    input_top_logprobs_val = input_top_logprobs_idx = (
-                        output_top_logprobs_val
-                    ) = output_top_logprobs_idx = None
-
-                # Compute the normalized logprobs for the requested tokens.
-                # Note that we pad a zero at the end for easy batching.
-                input_token_logprobs = all_logprobs[
-                    torch.arange(all_logprobs.shape[0], device="cuda"),
-                    torch.cat(
-                        [
-                            torch.cat(pruned_input_ids)[1:],
-                            torch.tensor([0], device="cuda"),
-                        ]
-                    ),
-                ]
-                normalized_prompt_logprobs = self._get_normalized_prompt_logprobs(
-                    input_token_logprobs,
-                    logits_metadata,
-                )
+                input_top_logprobs_val = input_top_logprobs_idx = None
+
+            # Compute the normalized logprobs for the requested tokens.
+            # Note that we pad a zero at the end for easy batching.
+            input_token_logprobs = input_logprobs[
+                torch.arange(input_logprobs.shape[0], device="cuda"),
+                torch.cat(
+                    [
+                        torch.cat(pruned_input_ids)[1:],
+                        torch.tensor([0], device="cuda"),
+                    ]
+                ),
+            ]
+            normalized_prompt_logprobs = self._get_normalized_prompt_logprobs(
+                input_token_logprobs,
+                logits_metadata,
+            )
 
-                return LogitsProcessorOutput(
-                    next_token_logits=last_logits,
-                    next_token_logprobs=last_logprobs,
-                    normalized_prompt_logprobs=normalized_prompt_logprobs,
-                    input_token_logprobs=input_token_logprobs,
-                    input_top_logprobs_val=input_top_logprobs_val,
-                    input_top_logprobs_idx=input_top_logprobs_idx,
-                    output_top_logprobs_val=output_top_logprobs_val,
-                    output_top_logprobs_idx=output_top_logprobs_idx,
-                )
+            return LogitsProcessorOutput(
+                next_token_logits=last_logits,
+                normalized_prompt_logprobs=normalized_prompt_logprobs,
+                input_token_logprobs=input_token_logprobs,
+                input_top_logprobs_val=input_top_logprobs_val,
+                input_top_logprobs_idx=input_top_logprobs_idx,
+            )
 
     def _get_logits(
         self,
@@ -269,9 +231,19 @@ class LogitsProcessor(nn.Module):
             # GGUF models
             logits = lm_head.linear_method.apply(lm_head, hidden_states, embedding_bias)
 
-        # Optional scaling factor
         if self.logit_scale is not None:
-            logits.mul_(self.logit_scale)  # In-place multiply
+            logits.mul_(self.logit_scale)
+
+        if self.do_tensor_parallel_all_gather:
+            logits = tensor_model_parallel_all_gather(logits)
+
+        # Compute the normalized logprobs for the requested tokens.
+        # Note that we pad a zero at the end for easy batching.
+        logits = logits[:, : self.config.vocab_size].float()
+
+        if self.final_logit_softcapping:
+            fused_softcap(logits, self.final_logit_softcapping)
+
         return logits
 
     @staticmethod
@@ -302,90 +274,73 @@ class LogitsProcessor(nn.Module):
         values = ret.values.tolist()
         indices = ret.indices.tolist()
 
-        if logits_metadata.forward_mode.is_decode():
-            output_top_logprobs_val = []
-            output_top_logprobs_idx = []
-            for i, k in enumerate(logits_metadata.top_logprobs_nums):
-                output_top_logprobs_val.append(values[i][:k])
-                output_top_logprobs_idx.append(indices[i][:k])
-            return None, None, output_top_logprobs_val, output_top_logprobs_idx
-        else:
-            input_top_logprobs_val, input_top_logprobs_idx = [], []
-            output_top_logprobs_val, output_top_logprobs_idx = [], []
+        input_top_logprobs_val, input_top_logprobs_idx = [], []
 
-            pt = 0
-            for k, pruned_len in zip(
-                logits_metadata.top_logprobs_nums,
-                logits_metadata.extend_logprob_pruned_lens_cpu,
-            ):
-                if pruned_len <= 0:
-                    input_top_logprobs_val.append([])
-                    input_top_logprobs_idx.append([])
-                    output_top_logprobs_val.append([])
-                    output_top_logprobs_idx.append([])
-                    continue
-
-                input_top_logprobs_val.append(
-                    [values[pt + j][:k] for j in range(pruned_len - 1)]
-                )
-                input_top_logprobs_idx.append(
-                    [indices[pt + j][:k] for j in range(pruned_len - 1)]
-                )
-                output_top_logprobs_val.append(
-                    list(
-                        values[pt + pruned_len - 1][:k],
-                    )
-                )
-                output_top_logprobs_idx.append(
-                    list(
-                        indices[pt + pruned_len - 1][:k],
-                    )
-                )
-                pt += pruned_len
+        pt = 0
+        for k, pruned_len in zip(
+            logits_metadata.top_logprobs_nums,
+            logits_metadata.extend_logprob_pruned_lens_cpu,
+        ):
+            if pruned_len <= 0:
+                input_top_logprobs_val.append([])
+                input_top_logprobs_idx.append([])
+                continue
 
-            return (
-                input_top_logprobs_val,
-                input_top_logprobs_idx,
-                output_top_logprobs_val,
-                output_top_logprobs_idx,
+            input_top_logprobs_val.append(
+                [values[pt + j][:k] for j in range(pruned_len - 1)]
             )
+            input_top_logprobs_idx.append(
+                [indices[pt + j][:k] for j in range(pruned_len - 1)]
+            )
+            pt += pruned_len
+
+        return input_top_logprobs_val, input_top_logprobs_idx
 
     @staticmethod
     def compute_temp_top_p_normalized_logprobs(
         last_logits: torch.Tensor, logits_metadata: LogitsMetadata
     ) -> torch.Tensor:
+        # TODO: Implement the temp and top-p normalization
         return torch.nn.functional.log_softmax(last_logits, dim=-1)
 
 
-def test():
-    all_logprobs = torch.tensor(
-        #       s                     s                s
-        [[0, 1, 2, 3], [1, 2, 3, 4], [2, 3, 4, 5], [3, 4, 5, 6], [4, 5, 6, 7]],
-        dtype=torch.float32,
-        device="cuda",
+@triton.jit
+def fused_softcap_kernel(
+    full_logits_ptr,
+    softcapping_value,
+    n_elements,
+    BLOCK_SIZE: tl.constexpr,
+):
+    pid = tl.program_id(0)
+    block_start = pid * BLOCK_SIZE
+    offsets = block_start + tl.arange(0, BLOCK_SIZE)
+    mask = offsets < n_elements
+
+    # Load values
+    x = tl.load(full_logits_ptr + offsets, mask=mask)
+
+    # Perform operations in-place
+    x = x / softcapping_value
+
+    # Manual tanh implementation using exp
+    exp2x = tl.exp(2 * x)
+    x = (exp2x - 1) / (exp2x + 1)
+
+    x = x * softcapping_value
+
+    # Store result
+    tl.store(full_logits_ptr + offsets, x, mask=mask)
+
+
+def fused_softcap(full_logits, final_logit_softcapping):
+    n_elements = full_logits.numel()
+    BLOCK_SIZE = 1024
+    grid = ((n_elements + BLOCK_SIZE - 1) // BLOCK_SIZE, 1, 1)
+
+    fused_softcap_kernel[grid](
+        full_logits_ptr=full_logits,
+        softcapping_value=final_logit_softcapping,
+        n_elements=n_elements,
+        BLOCK_SIZE=BLOCK_SIZE,
     )
-    seq_lens = torch.tensor([2, 0, 3, 0], dtype=torch.int32, device="cuda")
-    input_ids = torch.tensor([1, 2, 3, 0, 1], dtype=torch.int32, device="cuda")
-
-    token_logprobs = all_logprobs[
-        torch.arange(all_logprobs.shape[0], device="cuda"),
-        torch.cat([input_ids[1:], torch.tensor([0], device="cuda")]),
-    ]
-    logprobs_cumsum = torch.cumsum(token_logprobs, dim=0, dtype=torch.float32)
-
-    len_cumsum = torch.cumsum(seq_lens, dim=0)
-    start = torch.cat((torch.tensor([0], device="cuda"), len_cumsum[:-1]), 0)
-    end = start + seq_lens - 2
-    start.clamp_(min=0, max=token_logprobs.shape[0] - 1)
-    end.clamp_(min=0, max=token_logprobs.shape[0] - 1)
-    sum_logp = logprobs_cumsum[end] - logprobs_cumsum[start] + token_logprobs[start]
-
-    # assert logprobs == [2, _, 2, 4, _]
-    print("token logprobs", token_logprobs)
-    print("start", start)
-    print("end", end)
-    print("sum_logp", sum_logp)
-
-
-if __name__ == "__main__":
-    test()
+    return full_logits

sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}