[Model Runner V2] Optimize Gumbel Sampling Kernel (#29210)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

vllm/v1/worker/gpu/sampler.py

@@ -3,10 +3,9 @@
 from collections.abc import Callable
 
 import torch
-import triton
-import triton.language as tl
 
 from vllm.config.model import LogprobsMode
+from vllm.triton_utils import tl, triton
 from vllm.v1.outputs import LogprobsTensors, SamplerOutput
 from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p
 from vllm.v1.worker.gpu.states import SamplingMetadata
@@ -78,7 +77,10 @@ class Sampler:
 
 @triton.jit
 def _gumbel_sample_kernel(
-    sampled_ptr,
+    local_argmax_ptr,
+    local_argmax_stride,
+    local_max_ptr,
+    local_max_stride,
     logits_ptr,
     logits_stride,
     seeds_ptr,
@@ -88,14 +90,8 @@ def _gumbel_sample_kernel(
     BLOCK_SIZE: tl.constexpr,
 ):
     req_idx = tl.program_id(0)
-    is_greedy = tl.load(is_greedy_ptr + req_idx)
+    block_idx = tl.program_id(1)
-
+    block = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
-    if is_greedy:
-        # Greedy sampling. Don't apply gumbel noise.
-        max_val = float("-inf")
-        max_idx = 0
-        for i in range(0, vocab_size, BLOCK_SIZE):
-            block = i + tl.arange(0, BLOCK_SIZE)
     mask = block < vocab_size
     logits = tl.load(
         logits_ptr + req_idx * logits_stride + block,
@@ -103,42 +99,26 @@ def _gumbel_sample_kernel(
         other=float("-inf"),
     )
 
-            idx = tl.argmax(logits, axis=0)
+    is_greedy = tl.load(is_greedy_ptr + req_idx)
-            value = tl.max(logits, axis=0)
+    if not is_greedy:
-            is_greater = value > max_val
+        # Calculate the seed for gumbel noise.
-            max_val = tl.where(is_greater, value, max_val)
-            max_idx = tl.where(is_greater, i + idx, max_idx)
-        tl.store(sampled_ptr + req_idx, max_idx)
-        return
-
-    # Random sampling.
-    # Calculate gumbel seed.
         seed = tl.load(seeds_ptr + req_idx)
         pos = tl.load(pos_ptr + req_idx)
         gumbel_seed = tl.randint(seed, pos)
 
-    max_val = float("-inf")
-    max_idx = 0
-    for i in range(0, vocab_size, BLOCK_SIZE):
-        block = i + tl.arange(0, BLOCK_SIZE)
-        mask = block < vocab_size
-
         # Generate gumbel noise.
         r = tl.rand(gumbel_seed, block).to(tl.float64)
         gumbel_noise = -tl.log(-tl.log(r + 1e-20) + 1e-20)
         gumbel_noise = gumbel_noise.to(tl.float32)
 
         # Apply gumbel noise.
-        logits = tl.load(logits_ptr + req_idx * logits_stride + block, mask=mask)
         logits = tl.where(mask, logits + gumbel_noise, float("-inf"))
 
-        # Argmax to get the sampled token.
     idx = tl.argmax(logits, axis=0)
+    token_id = block_idx * BLOCK_SIZE + idx
     value = tl.max(logits, axis=0)
-        is_greater = value > max_val
+    tl.store(local_argmax_ptr + req_idx * local_argmax_stride + block_idx, token_id)
-        max_val = tl.where(is_greater, value, max_val)
+    tl.store(local_max_ptr + req_idx * local_max_stride + block_idx, value)
-        max_idx = tl.where(is_greater, i + idx, max_idx)
-    tl.store(sampled_ptr + req_idx, max_idx)
 
 
 def gumbel_sample(
@@ -148,23 +128,36 @@ def gumbel_sample(
     pos: torch.Tensor,  # [num_reqs]
 ) -> torch.Tensor:
     num_reqs, vocab_size = logits.shape
-    # NOTE(woosuk): Use int64 for later indexing.
+    BLOCK_SIZE = 1024
-    sampled = torch.empty(
+    num_blocks = triton.cdiv(vocab_size, BLOCK_SIZE)
+    local_argmax = torch.empty(
         num_reqs,
+        num_blocks,
         dtype=torch.int64,
         device=logits.device,
     )
-    _gumbel_sample_kernel[(num_reqs,)](
+    local_max = torch.empty(
-        sampled,
+        num_reqs,
+        num_blocks,
+        dtype=torch.float32,
+        device=logits.device,
+    )
+    _gumbel_sample_kernel[(num_reqs, num_blocks)](
+        local_argmax,
+        local_argmax.stride(0),
+        local_max,
+        local_max.stride(0),
         logits,
         logits.stride(0),
         seed,
         pos,
         is_greedy,
         vocab_size,
-        num_warps=8,
+        BLOCK_SIZE=BLOCK_SIZE,
-        BLOCK_SIZE=16384,  # type: ignore
     )
+    # NOTE(woosuk): Use int64 for later indexing.
+    max_block_idx = local_max.argmax(dim=-1, keepdim=True)
+    sampled = local_argmax.gather(dim=-1, index=max_block_idx).view(-1)
     return sampled