[MODELRUNNERV2]fix penality ops (#7013)

### What this PR does / why we need it?
fix penality ops for new version, and achieved a 10% performance
improvement

### How was this patch tested?
pytest
‎tests/e2e/nightly/single_node/ops/singlecard_ops/triton/test_penality.py
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2

Signed-off-by: shiyuan680 <917935075@qq.com>

tests/e2e/nightly/single_node/ops/singlecard_ops/triton/test_penality.py

@@ -1,63 +1,47 @@
 import pytest
 import torch
 
-from vllm_ascend.worker.v2.sample.penalties import apply_penalties_and_temperature
+from vllm_ascend.worker.v2.sample.penalties import apply_penalties
 
 DTYPES = [torch.bfloat16, torch.float16]
-NUM_REQS = [2, 4, 8]
+NUM_TOKENS = [2, 4, 8]
 VOCAB_SIZE = [151936]
 NUM_STATUS = [1, 4, 8, 16]
 SEEDS = [0]
 DEVICES = [f"npu:{0}"]
+NUM_SPECULATIVE_TOKENS = [0, 1, 3]
 DEFAULT_ATOL = 1e-3
 DEFAULT_RTOL = 1e-3
 
 
-class SamplingMetadata:
+def pytorch_apply_penalties(
-    def __init__(self,
+    logits: torch.Tensor,
+    idx_mapping: torch.Tensor,
+    token_ids: torch.Tensor,
+    expanded_local_pos: torch.Tensor,
     repetition_penalty: torch.Tensor,
     frequency_penalty: torch.Tensor,
     presence_penalty: torch.Tensor,
-                 temperature: torch.Tensor,
-                 idx_mapping: torch.Tensor,
     prompt_bin_mask: torch.Tensor,
-                 output_bin_counts: torch.Tensor):
+    output_bin_counts: torch.Tensor,
-        self.repetition_penalty = repetition_penalty
+    num_speculative_tokens: int,
-        self.frequency_penalty = frequency_penalty
-        self.presence_penalty = presence_penalty
-        self.temperature = temperature
-        self.idx_mapping = idx_mapping
-        self.prompt_bin_mask = prompt_bin_mask
-        self.output_bin_counts = output_bin_counts
-
-
-def pytorch_apply_penalties_and_temperature(
-    logits: torch.Tensor,
-    sampling_metadata: SamplingMetadata,
 ) -> torch.Tensor:
     """
     Pytorch equivalent implementation
     """
-    num_reqs, vocab_size = logits.shape
+    num_tokens, vocab_size = logits.shape
     device = logits.device
     dtype = logits.dtype
 
     logits_float = logits.float()
 
-    repetition_penalty = sampling_metadata.repetition_penalty
-    frequency_penalty = sampling_metadata.frequency_penalty
-    presence_penalty = sampling_metadata.presence_penalty
-    temperature = sampling_metadata.temperature
-    idx_mapping = sampling_metadata.idx_mapping
-    prompt_bin_mask = sampling_metadata.prompt_bin_mask
-    output_bin_counts = sampling_metadata.output_bin_counts
-
-    temperature = torch.where(temperature == 0.0, torch.ones_like(temperature), temperature)
-
     num_status = prompt_bin_mask.shape[0]
     num_packed = prompt_bin_mask.shape[1]
 
-    prompt_masks_unpacked = torch.zeros(num_status, vocab_size, dtype=torch.bool, device=device)
+    prompt_masks_unpacked = torch.zeros(
+        num_status, vocab_size, dtype=torch.bool,
+        device=device
+    )
 
     for state_idx in range(num_status):
         for packed_idx in range(num_packed):
@@ -69,82 +53,99 @@ def pytorch_apply_penalties_and_temperature(
                 if (packed_val >> bit_pos) & 1:
                     prompt_masks_unpacked[state_idx, start_idx + bit_pos] = True
 
-    for batch_idx in range(num_reqs):
+    for token_idx in range(num_tokens):
-        req_state_idx = idx_mapping[batch_idx].item()
+        req_state_idx = idx_mapping[token_idx].item()
         
-        rep_penalty = repetition_penalty[batch_idx].item()
+        rep_penalty = repetition_penalty[req_state_idx].item()
-        freq_penalty = frequency_penalty[batch_idx].item()
+        freq_penalty = frequency_penalty[req_state_idx].item()
-        pres_penalty = presence_penalty[batch_idx].item()
+        pres_penalty = presence_penalty[req_state_idx].item()
-        temp = temperature[batch_idx].item()
 
         use_rep_penalty = rep_penalty != 1.0
         use_freq_penalty = freq_penalty != 0.0
         use_pres_penalty = pres_penalty != 0.0
-        use_penalty = (use_rep_penalty or use_freq_penalty) or use_pres_penalty
+        use_penalty = use_rep_penalty or use_freq_penalty or use_pres_penalty
-        use_temperature = temp != 1.0
 
-        if not (use_penalty or use_temperature):
+        if not use_penalty:
             continue
 
         current_prompt_mask = prompt_masks_unpacked[req_state_idx]
-        current_output_counts = output_bin_counts[req_state_idx]
+        base_output_counts = output_bin_counts[req_state_idx]
-        output_bin_mask = current_output_counts > 0
+        
+        # Compute cumulative draft counts
+        pos = expanded_local_pos[token_idx].item()
+        start_idx_in_batch = token_idx - pos
+        draft_counts = torch.zeros(vocab_size, device=device, dtype=torch.int32)
+        
+        for prev_pos in range(num_speculative_tokens):
+            if prev_pos < pos:
+                prev_token = token_ids[start_idx_in_batch + prev_pos + 1].item()
+                draft_counts[prev_token] += 1
+        
+        # Total counts = base output counts + cumulative draft counts
+        total_output_counts = base_output_counts + draft_counts
+        output_bin_mask = total_output_counts > 0
 
         if use_rep_penalty:
             scale = torch.ones(vocab_size, device=device)
             mask = current_prompt_mask | output_bin_mask
             scale[mask] = rep_penalty
 
-            pos_mask = logits_float[batch_idx] > 0
+            pos_mask = logits_float[token_idx] > 0
             scale_factor = torch.where(pos_mask, 1.0 / scale, scale)
-            logits_float[batch_idx] *= scale_factor
+            logits_float[token_idx] *= scale_factor
 
         if use_freq_penalty:
-            logits_float[batch_idx] -= freq_penalty * current_output_counts.float()
+            logits_float[token_idx] -= freq_penalty * total_output_counts.float()
 
         if use_pres_penalty:
-            logits_float[batch_idx] -= pres_penalty * output_bin_mask.float()
+            logits_float[token_idx] -= pres_penalty * output_bin_mask.float()
-
-        if use_temperature:
-            logits_float[batch_idx] /= temp
 
     return logits_float.to(dtype)
 
 
 def create_test_data(
-    num_reqs: int = 8,
+    num_tokens: int = 8,
     vocab_size: int = 51200,
     num_status: int = 16,
+    num_speculative_tokens: int = 3,
     device: str = "npu",
     dtype: torch.dtype = torch.bfloat16,
     seed: int = 42,
 ):
-    """Create test data for penalties and temperature"""
+    """Create test data for penalties"""
     torch.manual_seed(seed)
 
-    logits = torch.randn(num_reqs, vocab_size, device=device, dtype=dtype)
+    logits = torch.randn(num_tokens, vocab_size, device=device, dtype=dtype)
 
-    repetition_penalty = torch.ones(num_reqs, device=device, dtype=torch.float32)
+    repetition_penalty = torch.ones(num_status, device=device, dtype=torch.float32)
-    for i in range(num_reqs):
+    for i in range(num_status):
         if torch.rand(1) > 0.3:
             repetition_penalty[i] = torch.rand(1, device=device).item() * 0.8 + 0.6
 
-    frequency_penalty = torch.zeros(num_reqs, device=device, dtype=torch.float32)
+    frequency_penalty = torch.zeros(num_status, device=device, dtype=torch.float32)
-    for i in range(num_reqs):
+    for i in range(num_status):
         if torch.rand(1) > 0.5:
             frequency_penalty[i] = torch.rand(1, device=device).item() * 0.2
 
-    presence_penalty = torch.zeros(num_reqs, device=device, dtype=torch.float32)
+    presence_penalty = torch.zeros(num_status, device=device, dtype=torch.float32)
-    for i in range(num_reqs):
+    for i in range(num_status):
         if torch.rand(1) > 0.5:
             presence_penalty[i] = torch.rand(1, device=device).item() * 0.2
 
-    temperature = torch.ones(num_reqs, device=device, dtype=torch.float32)
+    idx_mapping = torch.randint(
-    for i in range(num_reqs):
+        0, num_status, (num_tokens,), device=device,
-        if torch.rand(1) > 0.2:
+        dtype=torch.int32
-            presence_penalty[i] = torch.rand(1, device=device).item() * 1.8 + 0.2
+    )
     
-    idx_mapping = torch.randint(0, num_status, (num_reqs,), device=device, dtype=torch.int32)
+    # Create token_ids for speculative decoding
+    token_ids = torch.randint(0, vocab_size, (num_tokens,), device=device, dtype=torch.int32)
+    
+    # Create expanded_local_pos (position within speculative decoding window)
+    expanded_local_pos = torch.zeros(num_tokens, device=device, dtype=torch.int32)
+    for i in range(num_tokens):
+        expanded_local_pos[i] = torch.randint(
+            0, num_speculative_tokens + 1, (1,)
+        ).item()
 
     num_packed = (vocab_size + 31) // 32
     prompt_bin_mask = torch.zeros(num_status, num_packed, device=device, dtype=torch.int32)
@@ -161,44 +162,60 @@ def create_test_data(
     output_bin_counts = torch.zeros(num_status, vocab_size, device=device, dtype=torch.int32)
     for state_idx in range(num_status):
         num_output_tokens = max(1, vocab_size // 20)
-        output_tokens = torch.randint(0, vocab_size, (num_output_tokens, ))
+        output_tokens = torch.randint(0, vocab_size,
+                                      (num_output_tokens, ))
         counts = torch.randint(1, 10, (num_output_tokens,))
 
         for token, count in zip(output_tokens, counts):
             output_bin_counts[state_idx, token] = count
 
-    sampling_metadata = SamplingMetadata(
+    return (
-        repetition_penalty=repetition_penalty,
+        logits,
-        frequency_penalty=frequency_penalty,
+        idx_mapping,
-        presence_penalty=presence_penalty,
+        token_ids,
-        temperature=temperature,
+        expanded_local_pos,
-        idx_mapping=idx_mapping,
+        repetition_penalty,
-        prompt_bin_mask=prompt_bin_mask,
+        frequency_penalty,
-        output_bin_counts=output_bin_counts
+        presence_penalty,
+        prompt_bin_mask,
+        output_bin_counts,
+        num_speculative_tokens,
     )
 
-    return logits, sampling_metadata
 
-
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
-@pytest.mark.parametrize("num_reqs", NUM_REQS)
 @pytest.mark.parametrize("vocab_size", VOCAB_SIZE)
 @pytest.mark.parametrize("num_status", NUM_STATUS)
+@pytest.mark.parametrize("num_speculative_tokens", NUM_SPECULATIVE_TOKENS)
 @pytest.mark.parametrize("dtype", DTYPES)
 @pytest.mark.parametrize("seed", SEEDS)
 @pytest.mark.parametrize("device", DEVICES)
 @torch.inference_mode()
-def test_apply_penalties_and_temperature(
+def test_apply_penalties(
-    num_reqs,
+    num_tokens,
     vocab_size,
     num_status,
+    num_speculative_tokens,
     dtype,
     seed,
     device
 ):
-    logits_triton, sampling_metadata = create_test_data(
+    (
-        num_reqs=num_reqs,
+        logits_triton,
+        idx_mapping,
+        token_ids,
+        expanded_local_pos,
+        repetition_penalty,
+        frequency_penalty,
+        presence_penalty,
+        prompt_bin_mask,
+        output_bin_counts,
+        num_spec_tokens,
+    ) = create_test_data(
+        num_tokens=num_tokens,
         vocab_size=vocab_size,
         num_status=num_status,
+        num_speculative_tokens=num_speculative_tokens,
         device=device,
         dtype=dtype,
         seed=seed
@@ -206,10 +223,31 @@ def test_apply_penalties_and_temperature(
 
     logits_pytorch = logits_triton.clone()
 
-    apply_penalties_and_temperature(logits_triton, sampling_metadata)
+    apply_penalties(
+        logits_triton,
+        idx_mapping,
+        token_ids,
+        expanded_local_pos,
+        repetition_penalty,
+        frequency_penalty,
+        presence_penalty,
+        prompt_bin_mask,
+        output_bin_counts,
+        num_spec_tokens,
+    )
 
-    logits_pytorch_result = pytorch_apply_penalties_and_temperature(logits_pytorch,
+    logits_pytorch_result = pytorch_apply_penalties(
-                                                                    sampling_metadata)
+        logits_pytorch,
+        idx_mapping,
+        token_ids,
+        expanded_local_pos,
+        repetition_penalty,
+        frequency_penalty,
+        presence_penalty,
+        prompt_bin_mask,
+        output_bin_counts,
+        num_spec_tokens,
+    )
 
     atol = DEFAULT_ATOL
     rtol = DEFAULT_RTOL

vllm_ascend/worker/v2/sample/penalties.py

@@ -20,122 +20,141 @@
 
 import torch
 from vllm.triton_utils import tl, triton
-from vllm.v1.sample.metadata import SamplingMetadata
 
 
 @triton.jit
-def _penalties_and_temperature_kernel(
+def _penalties_kernel(
     logits_ptr,
     logits_stride,
-    repetition_penalty_ptr,
-    frequency_penalty_ptr,
-    presence_penalty_ptr,
-    temperature_ptr,
     idx_mapping_ptr,
+    token_ids_ptr,
+    expanded_local_pos_ptr,
+    penalties_ptr,
+    penalties_stride,
     prompt_bin_mask_ptr,
     prompt_bin_mask_stride,
     output_bin_counts_ptr,
     output_bin_counts_stride,
     vocab_size,
     BLOCK_SIZE: tl.constexpr,
+    INNER_BLOCK_SIZE: tl.constexpr,
+    MAX_SPEC_LEN: tl.constexpr,
 ):
-    batch_idx = tl.program_id(0)
+    token_idx = tl.program_id(0)
-    rep_penalty = tl.load(repetition_penalty_ptr + batch_idx)
+    req_state_idx = tl.load(idx_mapping_ptr + token_idx)
-    freq_penalty = tl.load(frequency_penalty_ptr + batch_idx)
+
-    pres_penalty = tl.load(presence_penalty_ptr + batch_idx)
+    # first load penalties once
-    temperature = tl.load(temperature_ptr + batch_idx)
+    rep_penalty = tl.load(penalties_ptr + req_state_idx * penalties_stride + 0)
-    temperature = tl.where(temperature == 0.0, 1.0, temperature)
+    freq_penalty = tl.load(penalties_ptr + req_state_idx * penalties_stride + 1)
+    pres_penalty = tl.load(penalties_ptr + req_state_idx * penalties_stride + 2)
 
     use_rep_penalty = rep_penalty != 1.0
     use_freq_penalty = freq_penalty != 0.0
     use_pres_penalty = pres_penalty != 0.0
-    # NOTE(Ronald1995): vllm original grammar `use_rep_penalty or
+
-    # use_freq_penalty or use_pres_penalty`,
+    # NPU doesn't support chained 'or' operations like 'A or B or C'
-    # change it to `(use_rep_penalty or use_freq_penalty) or use_pres_penalty`,
+    use_penalty = use_rep_penalty or use_freq_penalty
-    # because triton-ascend's compiler doesn't support chained boolean operator.
+    use_penalty = use_penalty or use_pres_penalty
-    use_penalty = (use_rep_penalty or use_freq_penalty) or use_pres_penalty
+
-    use_temperature = temperature != 1.0
+    if not use_penalty:
-    if not (use_penalty or use_temperature):
         # Early return to avoid loading logits.
         return
 
+    bit_masks = tl.full((INNER_BLOCK_SIZE // 32, 32), 1, dtype=tl.int32) << tl.arange(0, 32)
     block_idx = tl.program_id(1)
-    block = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    block_start = block_idx * BLOCK_SIZE
-    mask = block < vocab_size
+
-    logits = tl.load(logits_ptr + batch_idx * logits_stride + block, mask=mask)
+    pos = tl.load(expanded_local_pos_ptr + token_idx)
+    start_idx = token_idx - pos
+
+    inv_rep = 1.0 / rep_penalty
+
+    for inner_offset in tl.static_range(0, BLOCK_SIZE, INNER_BLOCK_SIZE):
+        inner_block_start = block_start + inner_offset
+        inner_block = inner_block_start + tl.arange(0, INNER_BLOCK_SIZE)
+        inner_mask = inner_block < vocab_size
+
+        logits = tl.load(logits_ptr + token_idx * logits_stride + inner_block, mask=inner_mask, other=0.0)
         logits = logits.to(tl.float32)
 
-    if use_penalty:
+        base_output_counts = tl.load(
-        req_state_idx = tl.load(idx_mapping_ptr + batch_idx)
+            output_bin_counts_ptr + req_state_idx * output_bin_counts_stride + inner_block,
-        output_bin_counts = tl.load(
+            mask=inner_mask,
-            output_bin_counts_ptr + req_state_idx * output_bin_counts_stride + block,
+            other=0,
-            mask=mask,
         )
-        # to use vector core, if use > 0 will use scalar to slow down performance
+
-        output_bin_mask = output_bin_counts != 0
+        # Compute cumulative draft_counts from previous positions in this request
+        total_counts = base_output_counts.to(tl.int32)
+        for prev_pos in tl.static_range(MAX_SPEC_LEN):
+            if prev_pos < pos:
+                load_idx = start_idx + prev_pos + 1
+                prev_token = tl.load(token_ids_ptr + load_idx)
+                total_counts += inner_block == prev_token
+
+        is_present = total_counts != 0
 
         # Apply repetition penalties.
         if use_rep_penalty:
-            packed_block = block_idx * BLOCK_SIZE // 32 + tl.arange(0, BLOCK_SIZE // 32)
+            packed_inner_block_start = inner_block_start // 32
-            packed_mask = tl.load(
+            packed_block = packed_inner_block_start + tl.arange(0, INNER_BLOCK_SIZE // 32)
+            valid_packed_mask = packed_block < tl.cdiv(vocab_size, 32)
+
+            packed_mask_val = tl.load(
                 prompt_bin_mask_ptr + req_state_idx * prompt_bin_mask_stride + packed_block,
-                mask=packed_block < tl.cdiv(vocab_size, 32),
+                mask=valid_packed_mask,
+                other=0,
             )
-            # the compiler itself does not optimize right-shift operations, so we change the same func
+            prompt_mask = ((packed_mask_val[:, None] & bit_masks) != 0).reshape(INNER_BLOCK_SIZE)
-            bit_masks = 1 << tl.arange(0, 32)
-            bit_masks_expanded = bit_masks[None, :]
 
-            packed_expanded = packed_mask[:, None]
+            needs_scaling = prompt_mask | is_present
-            bits_matrix = (packed_expanded & bit_masks_expanded) != 0
 
-            prompt_bin_mask = bits_matrix.reshape(BLOCK_SIZE)
+            base_factor = tl.where(logits > 0, inv_rep, rep_penalty)
+            logits = tl.where(needs_scaling, logits * base_factor, logits)
 
-            prompt_bin_mask = prompt_bin_mask.to(tl.int1)
+        freq_term = freq_penalty * total_counts.to(tl.float32)
-            prompt_bin_mask = prompt_bin_mask.reshape(BLOCK_SIZE)
+        pres_term = pres_penalty * is_present.to(tl.float32)
-
-            # If token appears in prompt or output, apply, otherwise use 1.0 for no-op.
-            scale = tl.where(prompt_bin_mask | output_bin_mask, rep_penalty, 1.0)
-            # If logits are positive, divide by penalty, otherwise multiply by penalty.
-            logits *= tl.where(logits > 0, 1.0 / scale, scale)
-
-        # Apply frequency penalties.
-        logits -= freq_penalty * output_bin_counts
-        # Apply presence penalties.
-        logits -= pres_penalty * output_bin_mask
-
-    # Apply temperature.
-    logits = logits / temperature
 
+        logits = logits - freq_term - pres_term
         # Store back to logits.
-    tl.store(logits_ptr + batch_idx * logits_stride + block, logits, mask=mask)
+        tl.store(logits_ptr + token_idx * logits_stride + inner_block, logits, mask=inner_mask)
 
 
-def apply_penalties_and_temperature(
+def apply_penalties(
     logits: torch.Tensor,
-    sampling_metadata: SamplingMetadata,
+    idx_mapping: torch.Tensor,
+    token_ids: torch.Tensor,
+    expanded_local_pos: torch.Tensor,
+    repetition_penalty: torch.Tensor,
+    frequency_penalty: torch.Tensor,
+    presence_penalty: torch.Tensor,
+    prompt_bin_mask: torch.Tensor,
+    output_bin_counts: torch.Tensor,
+    num_speculative_tokens: int,
 ) -> None:
-    """Override the function because there are some bugs
+    num_tokens, vocab_size = logits.shape
-    when _penalties_and_temperature_kernel runs on npu, we need to make some fixes.
+    BLOCK_SIZE = 8192
-    you could read NOTE(Ronald1995) comments to understand.
+    INNER_BLOCK_SIZE = 4096
-    """
-    num_reqs, vocab_size = logits.shape
-    # NOTE(Ronald1995): change BLOCK_SIZE from 8192 to 4096 in case UB overflow
-    # in triton-ascend.
-    BLOCK_SIZE = 4096
     num_blocks = triton.cdiv(vocab_size, BLOCK_SIZE)
-    # TODO(Ronald1995): Optimize the performance of the kernel in npu.
+
-    _penalties_and_temperature_kernel[(num_reqs, num_blocks)](
+    penalties = torch.stack(
+        [repetition_penalty[:num_tokens], frequency_penalty[:num_tokens], presence_penalty[:num_tokens]], dim=1
+    ).contiguous()
+    penalties_stride = penalties.stride(0)
+
+    _penalties_kernel[(num_tokens, num_blocks)](
         logits,
         logits.stride(0),
-        sampling_metadata.repetition_penalty,
+        idx_mapping,
-        sampling_metadata.frequency_penalty,
+        token_ids,
-        sampling_metadata.presence_penalty,
+        expanded_local_pos,
-        sampling_metadata.temperature,
+        penalties,
-        sampling_metadata.idx_mapping,
+        penalties_stride,
-        sampling_metadata.prompt_bin_mask,
+        prompt_bin_mask,
-        sampling_metadata.prompt_bin_mask.stride(0),
+        prompt_bin_mask.stride(0),
-        sampling_metadata.output_bin_counts,
+        output_bin_counts,
-        sampling_metadata.output_bin_counts.stride(0),
+        output_bin_counts.stride(0),
         vocab_size,
         BLOCK_SIZE=BLOCK_SIZE,
+        INNER_BLOCK_SIZE=INNER_BLOCK_SIZE,
+        MAX_SPEC_LEN=num_speculative_tokens,
     )