[V1][TPU] Enable Top K (vllm-project#15489)

Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Hyesoo Yang <hyeygit@gmail.com>
Co-authored-by: Hyesoo Yang <hyeygit@gmail.com>

Author: 2 people

tests/v1/tpu/test_sampler.py

@@ -1,4 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
+import random
+
 import pytest
 
 from vllm import LLM, envs
@@ -39,3 +41,19 @@ def test_sampler_different(model_name: str):
         # Unsupported `seed` param.
         sampling_params = SamplingParams(temperature=0.3, seed=42)
         output2 = llm.generate(prompts, sampling_params)
+
+    # Batch-case with TopK
+    for B in [4, 16]:
+        p = prompts * B
+        sampling_params = [
+            SamplingParams(
+                temperature=0.1,
+                min_p=0.8,
+                max_tokens=64,
+                # Vary number of ks
+                top_k=random.randint(4, 12)) for _ in range(B)
+        ]
+        # Make sure first two reqs have the same K
+        sampling_params[0] = sampling_params[1]
+        output = llm.generate(p, sampling_params)
+        assert output[0].outputs[0].text == output[1].outputs[0].text

tests/v1/tpu/test_topk_topp_sampler.py

@@ -5,7 +5,8 @@
 import torch
 
 from vllm.platforms import current_platform
-from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p_tpu
+from vllm.v1.sample.ops.topk_topp_sampler import (apply_top_k_top_p,
+                                                  apply_top_k_top_p_tpu)
 
 if not current_platform.is_tpu():
     pytest.skip("This test needs a TPU.", allow_module_level=True)
@@ -16,6 +17,25 @@
 TOLERANCE = 1e-6
 
 
+def test_topk_equivalence_to_native_impl():
+    with torch.device(xm.xla_device()):
+        xm.set_rng_state(seed=33)
+
+        logits = torch.rand((BATCH_SIZE, VOCAB_SIZE))
+
+        # Random top-k values between 1 and 10.
+        k = torch.randint(1, 10, (BATCH_SIZE, ))
+
+        # Set k=vocab_size for ~50% of requests in the batch (top-k disabled).
+        k.masked_fill_(torch.randint(0, 2, (BATCH_SIZE, ), dtype=bool),
+                       VOCAB_SIZE)
+
+        result_tpu = apply_top_k_top_p_tpu(logits=logits.clone(), k=k, p=None)
+
+        result_native = apply_top_k_top_p(logits=logits.clone(), k=k, p=None)
+        assert torch.allclose(result_native, result_tpu)
+
+
 def test_topp_result_sums_past_p():
     with torch.device(xm.xla_device()):
         xm.set_rng_state(seed=33)

vllm/envs.py

@@ -103,7 +103,6 @@
     VLLM_DP_MASTER_PORT: int = 0
     VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
     VLLM_V0_USE_OUTLINES_CACHE: bool = False
-    VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION: bool = False
     VLLM_TPU_BUCKET_PADDING_GAP: int = 0
     VLLM_USE_DEEP_GEMM: bool = False
     VLLM_XGRAMMAR_CACHE_MB: int = 0
@@ -685,11 +684,6 @@ def maybe_convert_int(value: Optional[str]) -> Optional[int]:
     "VLLM_V0_USE_OUTLINES_CACHE":
     lambda: os.environ.get("VLLM_V0_USE_OUTLINES_CACHE", "0") == "1",
 
-    # If set, disables TPU-specific optimization for top-k & top-p sampling
-    "VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION":
-    lambda: bool(int(os.environ["VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION"]))
-    if "VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION" in os.environ else None,
-
     # Gap between padding buckets for the forward pass. So we have
     # 8, we will run forward pass with [16, 24, 32, ...].
     "VLLM_TPU_BUCKET_PADDING_GAP":

vllm/v1/sample/ops/topk_topp_sampler.py

@@ -72,14 +72,7 @@ def __init__(self):
                     "best performance, please install FlashInfer.")
                 self.forward = self.forward_native
         elif current_platform.is_tpu():
-            if envs.VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION:
-                logger.warning(
-                    "TPU-specific optimization for top-k & top-p sampling are "
-                    "disabled, falling back to PyTorch-native implementation "
-                    "which could be very slow.")
-                self.forward = self.forward_native
-            else:
-                self.forward = self.forward_tpu
+            self.forward = self.forward_tpu
         else:
             self.forward = self.forward_native
 
@@ -146,12 +139,22 @@ def apply_top_k_top_p_tpu(
     chance of being chosen during final sampling, so we can consider the tie
     being broken then.
     """
+    probs = logits.softmax(dim=-1)
+    probs_sort, _ = probs.sort(dim=-1, descending=False)
+
     if k is not None:
-        logits = apply_top_k_only(logits, k)
+        top_k_count = probs_sort.size(1) - k.to(torch.long)  # shape: (batch, )
+        top_k_count = top_k_count.unsqueeze(dim=1)
+        top_k_cutoff = probs_sort.gather(-1, top_k_count)
+
+        # Make sure the no top-k rows are no-op.
+        no_top_k_mask = (k == logits.shape[1]).unsqueeze(dim=1)
+        top_k_cutoff.masked_fill_(no_top_k_mask, -float("inf"))
+
+        elements_to_discard = probs < top_k_cutoff
+        logits.masked_fill_(elements_to_discard, -float("inf"))
 
     if p is not None:
-        probs = logits.softmax(dim=-1)
-        probs_sort, _ = probs.sort(dim=-1, descending=False)
         cumprob = torch.cumsum(probs_sort, dim=-1)
         top_p_mask = cumprob <= 1 - p.unsqueeze(dim=1)
         top_p_mask[:, -1] = False  # at least one
@@ -224,7 +227,7 @@ def apply_top_k_only(
     max_top_k = k.max()
     # topk.values tensor has shape [batch_size, max_top_k].
     # Convert top k to 0-based index in range [0, max_top_k).
-    k_index = k.sub_(1).unsqueeze(1).expand(logits.shape[0], 1)
+    k_index = k.sub_(1).unsqueeze(1)
     top_k_mask = logits.topk(max_top_k, dim=1).values.gather(1, k_index.long())
     # Handle non-topk rows.
     top_k_mask.masked_fill_(no_top_k_mask.unsqueeze(1), -float("inf"))

vllm/v1/sample/tpu/metadata.py

@@ -10,7 +10,7 @@
     temperature=-1.0,
     min_p=0.0,
     # strictly disabled for now
-    # top_k=-1,
+    top_k=0,
     # top_p=0.0,
     # frequency_penalties=0.0,
     # presence_penalties=0.0,
@@ -99,11 +99,13 @@ def fill_slice(cpu_tensor: torch.Tensor, fill_val) -> torch.Tensor:
 
         fill_slice(input_batch.temperature_cpu_tensor,
                    DEFAULT_SAMPLING_PARAMS["temperature"])
-        # TODO Temporarily disabled until sampling options are enabled
-        # fill_slice(input_batch.top_p_cpu_tensor)
-        # fill_slice(input_batch.top_k_cpu_tensor)
         fill_slice(input_batch.min_p_cpu_tensor,
                    DEFAULT_SAMPLING_PARAMS["min_p"])
+        fill_slice(input_batch.top_k_cpu_tensor,
+                   DEFAULT_SAMPLING_PARAMS["top_k"])
+        # TODO Temporarily disabled until sampling options are enabled
+        # fill_slice(input_batch.top_p_cpu_tensor,
+        #            DEFAULT_SAMPLING_PARAMS["top_p"])
 
         # Slice persistent device tensors to a fixed pre-compiled padded shape.
         return cls(
@@ -112,6 +114,7 @@ def fill_slice(cpu_tensor: torch.Tensor, fill_val) -> torch.Tensor:
             all_greedy=input_batch.all_greedy,
             # TODO enable more and avoid returning None values
             top_p=None,  # input_batch.top_p[:padded_num_reqs],
-            top_k=None,  # input_batch.top_k[:padded_num_reqs],
+            top_k=input_batch.top_k_cpu_tensor[:padded_num_reqs].to(
+                xla_device),
             min_p=input_batch.min_p_cpu_tensor[:padded_num_reqs].to(
                 xla_device))

vllm/v1/worker/tpu_model_runner.py

@@ -920,14 +920,19 @@ def _precompile_select_hidden_states(self) -> None:
                                       device=self.device)
                 torch._dynamo.mark_dynamic(indices, 0)
                 self.select_hidden_states(dummy_hidden, indices)
-            logger.info("  -- num_tokens: %d", num_tokens)
+                logger.info("  -- num_tokens: %d, num_seqs: %d", num_tokens,
+                            num_reqs)
+                # Requests can't be more than tokens. But do compile for the
+                # next bigger value in case num_tokens uses bucketed padding.
+                if num_reqs >= min(num_tokens, self.max_num_reqs):
+                    break
         xm.wait_device_ops()
         end = time.perf_counter()
         logger.info("Compilation finished in in %.2f [secs].", end - start)
         self._update_num_xla_graphs("select_hidden_states")
 
     def _precompile_sample_from_hidden(self) -> None:
-        logger.info("Compiling sampling with different input shapes.")
+        logger.info("Compiling sampling with different num_reqs.")
         start = time.perf_counter()
         hsize = self.model_config.get_hidden_size()
         for num_reqs in self.num_reqs_paddings: