[TPU][V1] Add support for top-logprobs (vllm-project#17072)

Signed-off-by: NickLucche <[email protected]>

Author: NickLucche

tests/v1/tpu/test_sampler.py

@@ -61,3 +61,51 @@ def test_sampler_different(model_name: str):
         # to have deterministic results over many tokens, tests the first ~20
         # tokens match.
         assert output[0].outputs[0].text[:20] == output[1].outputs[0].text[:20]
+
+
+@pytest.mark.parametrize("model_name", ["Qwen/Qwen2.5-1.5B-Instruct"])
+# TODO TPU will appear busy if we fan-out test params here
+@pytest.mark.parametrize("n_prompts", [1])
+@pytest.mark.skipif(not current_platform.is_tpu(),
+                    reason="This test needs a TPU")
+def test_logprobs(model_name: str, n_prompts: int):
+    """
+    Request top logprobs with different sampling settings and check
+    that results contains the requested number, ordered ascendingly.  
+    """
+
+    def check_num_logprobs(logprobs, expected_num: int):
+        for step in logprobs:
+            prev_logp = 1.0
+            # order by rank
+            sorted_step = dict(
+                sorted(step.items(), key=lambda item: item[1].rank))
+
+            # Can contain the sampled token
+            assert len(step) == expected_num or len(step) == expected_num + 1
+            # Check results are ordered by prob value
+            for rankno, (tid, logp) in enumerate(sorted_step.items()):
+                assert logp.logprob <= prev_logp
+                prev_logp = logp.logprob
+                assert logp.rank == rankno + 1
+
+    llm = LLM(model_name,
+              enforce_eager=False,
+              max_num_seqs=1,
+              max_model_len=128,
+              max_num_batched_tokens=128)
+    prompts = [
+        "Write a short story about a robot that dreams for the first time."
+    ] * n_prompts
+    greedy_sampling_params = SamplingParams(temperature=0.0, max_tokens=64,\
+         logprobs=4)
+    regular_sampling_params = SamplingParams(temperature=0.4, max_tokens=64,\
+         logprobs=4)
+    topkp_sampling_params = SamplingParams(temperature=0.4, max_tokens=64,\
+         logprobs=4, top_k=12, top_p=0.5)
+
+    for sp in [greedy_sampling_params, regular_sampling_params, \
+               topkp_sampling_params]:
+        output = llm.generate(prompts, sp)
+        for o in output:
+            check_num_logprobs(o.outputs[0].logprobs, 4)

vllm/v1/sample/tpu/metadata.py

@@ -31,8 +31,10 @@ class TPUSupportedSamplingMetadata:
 
     all_greedy: bool = True
 
-    # unsupported, you need to return an extra tensor of static size BxV
-    max_num_logprobs = None
+    # Whether logprobs are to be gathered in this batch of request. To balance
+    # out compile time and runtime, a fixed `max_number_logprobs` value is used
+    # when gathering logprobs, regardless of the values specified in the batch.
+    logprobs: bool = False
 
     # TODO No penalties for now
     no_penalties: bool = True
@@ -84,10 +86,12 @@ def from_input_batch(
                 we want to pre-compile a graph with sampling parameters, even if
                 they are not strictly needed for greedy decoding.
         """
+        needs_logprobs = input_batch.max_num_logprobs>0 if \
+            input_batch.max_num_logprobs else False
         # Early return to avoid unnecessary cpu to tpu copy
         if (input_batch.all_greedy is True
                 and generate_params_if_all_greedy is False):
-            return cls(all_greedy=True)
+            return cls(all_greedy=True, logprobs=needs_logprobs)
 
         num_reqs = input_batch.num_reqs
 
@@ -115,4 +119,5 @@ def fill_slice(cpu_tensor: torch.Tensor, fill_val) -> torch.Tensor:
             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))
+                xla_device),
+            logprobs=needs_logprobs)

vllm/v1/sample/tpu/sampler.py

@@ -22,35 +22,25 @@ def forward(
         logits: torch.Tensor,
         sampling_metadata: TPUSupportedSamplingMetadata,
     ) -> SamplerOutput:
-        # NOTE(woosuk): Use the original logits (before any penalties or
-        # temperature scaling) for the top-k logprobs.
-        # This is different from the V0 sampler, which uses the logits that
-        # is used for sampling (after penalties and temperature scaling).
-
         # Use float32 for the logits.
         logits = logits.to(torch.float32)
         # Sample the next token.
         sampled = self.sample(logits, sampling_metadata)
 
-        # Use int32 to reduce the tensor size.
-        sampled = sampled.to(torch.int32)
-
-        # These are GPU tensors.
+        # These are TPU tensors.
         sampler_output = SamplerOutput(
             # The sampled tokens are expanded to 2D tensor with shape
             # [num_requests, 1], where each row represents one generated
             # token per request.
             sampled_token_ids=sampled.unsqueeze(-1),
-            logprobs_tensors=None,
-        )
+            logprobs_tensors=None)
         return sampler_output
 
     def apply_temperature(
         self,
         logits: torch.Tensor,
         temp: torch.Tensor,
     ) -> torch.Tensor:
-        # Use in-place division to avoid creating a new tensor.
         return logits.div_(temp.unsqueeze(dim=1))
 
     def greedy_sample(self, logits: torch.Tensor) -> torch.Tensor:

vllm/v1/worker/tpu_model_runner.py

@@ -791,8 +791,18 @@ def execute_model(
                                             arange)
         selected_token_ids = self.sample_from_logits(logits,
                                                      tpu_sampling_metadata)
+
+        # NOTE (NickLucche) Use the original logits (before any penalties or
+        # temperature scaling) for the top-k logprobs. We can't enforce it due
+        # to recompilations outside torch.compiled code, so just make sure
+        # `sample_from_logits` does not modify the logits in-place.
+        logprobs = self.gather_logprobs(logits, selected_token_ids) \
+            if tpu_sampling_metadata.logprobs else None
+
         # Remove padding on cpu and keep dynamic op outside of xla graph.
         selected_token_ids = selected_token_ids.cpu()[:num_reqs]
+        logprobs_lists = logprobs.tolists() \
+            if tpu_sampling_metadata.logprobs else None
 
         # Update the cache state concurrently. Code above will not block until
         # we use `selected_token_ids`. Add mark_step if post-processing changes
@@ -862,7 +872,7 @@ def execute_model(
             req_id_to_index=self.input_batch.req_id_to_index,
             sampled_token_ids=valid_sampled_token_ids,
             spec_token_ids=None,
-            logprobs=None,
+            logprobs=logprobs_lists,
             prompt_logprobs_dict=prompt_logprobs_dict,
         )
 
@@ -1121,6 +1131,22 @@ def _precompile_sample_from_logits(self) -> None:
         logger.info("Compilation finished in %.2f [secs].", end - start)
         self._update_num_xla_graphs("sample_from_logits")
 
+    def _precompile_gather_logprobs(self) -> None:
+        logger.info("Compiling gather_logprobs with different input shapes.")
+        start = time.perf_counter()
+        for num_reqs in self.num_reqs_paddings:
+            dummy_logits = torch.zeros((num_reqs, self.vocab_size),
+                                       device=self.device,
+                                       dtype=self._hidden_states_dtype)
+            dummy_tokens = torch.zeros((num_reqs, 1),
+                                       dtype=torch.int64).to(self.device)
+            self.gather_logprobs(dummy_logits, dummy_tokens)
+            logger.info("  -- num_seqs: %d", num_reqs)
+        xm.wait_device_ops()
+        end = time.perf_counter()
+        logger.info("Compilation finished in %.2f [secs].", end - start)
+        self._update_num_xla_graphs("gather_logprobs")
+
     def capture_model(self) -> None:
         """
         Precompile all the subgraphs with possible input shapes.
@@ -1131,6 +1157,7 @@ def capture_model(self) -> None:
         self._precompile_compute_logits()
         self._precompile_structured_decoding()
         self._precompile_sample_from_logits()
+        self._precompile_gather_logprobs()
 
     def profile_run(
         self,
@@ -1254,13 +1281,31 @@ def compute_logits(self,
     def sample_from_logits(
             self, logits: torch.Tensor,
             sampling_metadata: TPUSupportedSamplingMetadata) -> torch.Tensor:
+        """
+        Sample with xla-friendly function. This function is to be traced 
+        separately from `forward` for lighter compilation overhead.
+        """
         if sampling_metadata.all_greedy:
             out_tokens = torch.argmax(logits, dim=-1, keepdim=True)
         else:
             out_tokens = self.sampler(logits,
                                       sampling_metadata).sampled_token_ids
         return out_tokens
 
+    @torch.compile(backend="openxla", fullgraph=True, dynamic=False)
+    def gather_logprobs(self, logits: torch.Tensor,
+                        sampled_tokens: torch.Tensor) -> LogprobsTensors:
+        """
+        Gather the top_logprobs with corresponding tokens. Use a fixed number
+        of logprobs as an alternative to having multiple pre-compiled graphs.
+        Select the number of logprobs actually demanded by each request on CPU.
+        """
+        logprobs = self.sampler.compute_logprobs(logits)
+        return self.sampler.gather_logprobs(
+            logprobs,
+            self.model_config.max_logprobs,
+            token_ids=sampled_tokens.squeeze(-1))
+
     @torch.compile(backend="openxla", fullgraph=True, dynamic=False)
     def structured_decode(self, require_struct_decoding: torch.Tensor,
                           grammar_bitmask: torch.Tensor, logits: torch.Tensor,