[Feature] KV cache per-token-head INT8/FP8 quantization (#38378)

Signed-off-by: JartX <sagformas@epdcenter.es>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: yangyang4991 <yangyang4991@gmail.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Isotr0py <2037008807@qq.com>

docs/design/attention_backends.md

@@ -177,7 +177,7 @@ Priority is **1 = highest** (tried first).
 | `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto` | %16 | Any | ✅ | ✅ | ❌ | All | N/A |
 | `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ✅ | ❌ | Decoder, Encoder, Encoder Only | N/A |
 | `TREE_ATTN` | | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
-| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
+| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
 
 > **†** FlashInfer uses TRTLLM attention on Blackwell (SM100), which supports sinks. Disable via `--attention-config.use_trtllm_attention=0`.
 >

tests/models/quantization/test_per_token_kv_cache.py

@@ -0,0 +1,94 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""End-to-end accuracy tests for per-token-head KV cache quantization.
+
+Compares logprobs between a baseline bf16 model and the same model with
+per-token-head quantized KV cache (int8 or fp8) using the Triton attention
+backend.
+
+Run: pytest tests/models/quantization/test_per_token_kv_cache.py -v -s
+"""
+
+import pytest
+
+from vllm.platforms import current_platform
+
+from ..utils import check_logprobs_close
+
+
+@pytest.mark.skipif(
+    not current_platform.is_cuda_alike(),
+    reason="Per-token-head KV cache requires CUDA or ROCm GPU.",
+)
+@pytest.mark.parametrize(
+    "base_model,test_model",
+    [
+        (
+            "meta-llama/Llama-3.2-1B-Instruct",
+            "meta-llama/Llama-3.2-1B-Instruct",
+        ),
+    ],
+)
+@pytest.mark.parametrize(
+    "kv_cache_dtype", ["int8_per_token_head", "fp8_per_token_head"]
+)
+@pytest.mark.parametrize("max_tokens", [4])
+@pytest.mark.parametrize("enforce_eager", [True])
+@pytest.mark.parametrize("backend", ["TRITON_ATTN"])
+@pytest.mark.parametrize("tensor_parallel_size", [1])
+def test_per_token_head_kv_cache_accuracy(
+    vllm_runner,
+    example_prompts,
+    base_model: str,
+    test_model: str,
+    kv_cache_dtype: str,
+    max_tokens: int,
+    enforce_eager: bool,
+    backend: str,
+    tensor_parallel_size: int,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Compare logprobs between bf16 baseline and per-token-head quantized KV
+    cache.
+
+    Uses calculate_kv_scales (dynamic scale computation) since there are
+    no per-token-head calibrated checkpoints available yet.
+    """
+    with monkeypatch.context() as m:
+        m.setenv("TOKENIZERS_PARALLELISM", "true")
+
+        MAX_MODEL_LEN = 1024
+        NUM_LOG_PROBS = 8
+
+        with vllm_runner(
+            base_model,
+            max_model_len=MAX_MODEL_LEN,
+            tensor_parallel_size=tensor_parallel_size,
+            enforce_eager=enforce_eager,
+            kv_cache_dtype="auto",
+            attention_config={"backend": backend},
+        ) as vllm_model:
+            baseline_outputs = vllm_model.generate_greedy_logprobs(
+                example_prompts, max_tokens, NUM_LOG_PROBS
+            )
+
+        with vllm_runner(
+            test_model,
+            max_model_len=MAX_MODEL_LEN,
+            tensor_parallel_size=tensor_parallel_size,
+            enforce_eager=enforce_eager,
+            kv_cache_dtype=kv_cache_dtype,
+            calculate_kv_scales=True,
+            attention_config={"backend": backend},
+        ) as vllm_model:
+            test_outputs = vllm_model.generate_greedy_logprobs(
+                example_prompts, max_tokens, NUM_LOG_PROBS
+            )
+
+        check_logprobs_close(
+            outputs_0_lst=baseline_outputs,
+            outputs_1_lst=test_outputs,
+            name_0="bf16_kv_cache",
+            name_1=f"{kv_cache_dtype}_kv_cache",
+        )

tests/quantization/test_per_token_kv_cache.py

@@ -0,0 +1,560 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for per-token-head KV cache quantization (INT8 and FP8).
+
+Covers:
+- Per-token-head Triton reshape-and-cache kernel
+- Round-trip quantize/dequantize accuracy
+- process_weights_after_loading early-return path
+- End-to-end integration with Triton unified attention kernel
+
+Run: pytest tests/quantization/test_per_token_kv_cache.py -v -s
+"""
+
+import random
+from dataclasses import dataclass
+from unittest.mock import MagicMock
+
+import pytest
+import torch
+
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    get_fp8_min_max,
+)
+from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
+from vllm.v1.kv_cache_interface import KVQuantMode, is_quantized_kv_cache
+
+# Skip entire module if no CUDA/ROCm GPU available
+pytestmark = [
+    pytest.mark.skipif(
+        not current_platform.is_cuda_alike(),
+        reason="Per-token-head KV cache tests require CUDA or ROCm GPU.",
+    ),
+]
+
+# ---------------------------------------------------------------------------
+# Test parameters
+# ---------------------------------------------------------------------------
+NUM_TOKENS = [1, 7, 42]
+NUM_KV_HEADS = [1, 4, 8]
+HEAD_SIZES = [64, 128]
+BLOCK_SIZES = [16]
+SEEDS = [0]
+
+# Platform-dependent FP8 dtype and range
+FP8_DTYPE = current_platform.fp8_dtype()
+FP8_MIN, FP8_MAX = get_fp8_min_max()
+
+
+# ---------------------------------------------------------------------------
+# Per-dtype quantization config
+# ---------------------------------------------------------------------------
+@dataclass(frozen=True)
+class QuantConfig:
+    """Quantization parameters for a given cache dtype."""
+
+    cache_dtype: torch.dtype  # torch.int8 or FP8_DTYPE
+    kv_cache_dtype_str: str  # "int8_per_token_head" or "fp8_per_token_head"
+    quant_max: float
+    quant_min: float
+    kv_quant_mode: KVQuantMode
+    # INT8 Triton stores truncate; FP8 hardware casts round.
+    uses_trunc: bool
+
+
+INT8_CONFIG = QuantConfig(
+    cache_dtype=torch.int8,
+    kv_cache_dtype_str="int8_per_token_head",
+    quant_max=127.0,
+    quant_min=-128.0,
+    kv_quant_mode=KVQuantMode.INT8_PER_TOKEN_HEAD,
+    uses_trunc=True,
+)
+FP8_CONFIG = QuantConfig(
+    cache_dtype=FP8_DTYPE,
+    kv_cache_dtype_str="fp8_per_token_head",
+    quant_max=FP8_MAX,
+    quant_min=FP8_MIN,
+    kv_quant_mode=KVQuantMode.FP8_PER_TOKEN_HEAD,
+    uses_trunc=False,
+)
+
+QUANT_CONFIGS = [INT8_CONFIG, FP8_CONFIG]
+
+
+@pytest.fixture(params=QUANT_CONFIGS, ids=["int8", "fp8"])
+def qcfg(request) -> QuantConfig:
+    return request.param
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+def _quantize_per_token_head_ref(
+    data: torch.Tensor,  # [num_tokens, num_heads, head_size]
+    cfg: QuantConfig,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Reference per-token-head quantization (one scale per token per head).
+
+    Returns (quantized, scales) where scales is [num_tokens, num_heads].
+    """
+    absmax = data.float().abs().amax(dim=2)  # [num_tokens, num_heads]
+    scales = (absmax / cfg.quant_max).clamp(min=1e-6)
+    scaled = data.float() * (1.0 / scales[:, :, None])
+    if cfg.uses_trunc:
+        q = scaled.round().clamp(cfg.quant_min, cfg.quant_max).to(cfg.cache_dtype)
+    else:
+        q = scaled.clamp(cfg.quant_min, cfg.quant_max).to(cfg.cache_dtype)
+    return q, scales
+
+
+# ===========================================================================
+# 1. is_quantized_kv_cache / get_kv_quant_mode
+# ===========================================================================
+class TestIsQuantizedKvCache:
+    def test_fp8_variants(self):
+        assert is_quantized_kv_cache("fp8")
+        assert is_quantized_kv_cache("fp8_e4m3")
+        assert is_quantized_kv_cache("fp8_e5m2")
+
+    def test_int8_per_token_head(self):
+        assert is_quantized_kv_cache("int8_per_token_head")
+
+    def test_fp8_per_token_head(self):
+        assert is_quantized_kv_cache("fp8_per_token_head")
+
+    def test_auto(self):
+        assert not is_quantized_kv_cache("auto")
+
+    def test_bfloat16(self):
+        assert not is_quantized_kv_cache("bfloat16")
+
+    def test_kv_quant_mode_int8(self):
+        from vllm.v1.kv_cache_interface import get_kv_quant_mode
+
+        assert (
+            get_kv_quant_mode("int8_per_token_head") == KVQuantMode.INT8_PER_TOKEN_HEAD
+        )
+
+    def test_kv_quant_mode_fp8(self):
+        from vllm.v1.kv_cache_interface import get_kv_quant_mode
+
+        assert get_kv_quant_mode("fp8_per_token_head") == KVQuantMode.FP8_PER_TOKEN_HEAD
+
+
+# ===========================================================================
+# 2. Triton per-token-head kernel (reshape-and-cache)
+# ===========================================================================
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("num_heads", NUM_KV_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_reshape_and_cache_per_token_head(
+    qcfg: QuantConfig,
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+    seed: int,
+):
+    """Test triton_reshape_and_cache_flash_per_token_head_quant kernel."""
+    from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
+        triton_reshape_and_cache_flash_per_token_head_quant,
+    )
+
+    set_random_seed(seed)
+    torch.set_default_device("cuda")
+
+    num_blocks = (num_tokens + block_size - 1) // block_size + 4
+
+    key = torch.randn(num_tokens, num_heads, head_size, dtype=torch.bfloat16)
+    value = torch.randn(num_tokens, num_heads, head_size, dtype=torch.bfloat16)
+
+    key_cache = torch.zeros(
+        num_blocks, block_size, num_heads, head_size, dtype=qcfg.cache_dtype
+    )
+    value_cache = torch.zeros(
+        num_blocks, block_size, num_heads, head_size, dtype=qcfg.cache_dtype
+    )
+    k_scale_cache = torch.ones(num_blocks, block_size, num_heads, dtype=torch.float32)
+    v_scale_cache = torch.ones(num_blocks, block_size, num_heads, dtype=torch.float32)
+
+    num_slots = block_size * num_blocks
+    slot_mapping = torch.tensor(
+        random.sample(range(num_slots), num_tokens), dtype=torch.long
+    )
+
+    triton_reshape_and_cache_flash_per_token_head_quant(
+        key,
+        value,
+        key_cache,
+        value_cache,
+        k_scale_cache,
+        v_scale_cache,
+        slot_mapping,
+    )
+
+    # Reference
+    ref_k_quant, ref_k_scales = _quantize_per_token_head_ref(key, qcfg)
+    ref_v_quant, ref_v_scales = _quantize_per_token_head_ref(value, qcfg)
+
+    # Compare dequantized values rather than raw quantized values.
+    # Triton and PyTorch reductions can differ at FP8 rounding boundaries
+    # (up to 32 in quantized domain for fp8_e4m3), but the dequantized
+    # error is bounded by the scale.
+    for i, slot in enumerate(slot_mapping.tolist()):
+        blk = slot // block_size
+        off = slot % block_size
+
+        actual_k_scale = k_scale_cache[blk, off]  # [num_heads]
+        k_deq = key_cache[blk, off].float() * actual_k_scale[:, None]
+        k_ref_deq = key[i].float()
+        torch.testing.assert_close(
+            k_deq,
+            k_ref_deq,
+            atol=0.1,
+            rtol=0.1,
+        )
+        actual_v_scale = v_scale_cache[blk, off]  # [num_heads]
+        v_deq = value_cache[blk, off].float() * actual_v_scale[:, None]
+        v_ref_deq = value[i].float()
+        torch.testing.assert_close(
+            v_deq,
+            v_ref_deq,
+            atol=0.1,
+            rtol=0.1,
+        )
+        # Per-head scales: [num_heads]
+        torch.testing.assert_close(
+            k_scale_cache[blk, off], ref_k_scales[i], atol=1e-4, rtol=1e-3
+        )
+        torch.testing.assert_close(
+            v_scale_cache[blk, off], ref_v_scales[i], atol=1e-4, rtol=1e-3
+        )
+
+
+# ===========================================================================
+# 3. Per-token-head round-trip accuracy (quantize -> dequantize)
+# ===========================================================================
+@pytest.mark.parametrize("num_tokens", [1, 16])
+@pytest.mark.parametrize("num_heads", [4])
+@pytest.mark.parametrize("head_size", [128])
+@pytest.mark.parametrize("block_size", [16])
+@torch.inference_mode()
+def test_per_token_head_round_trip_accuracy(
+    qcfg: QuantConfig,
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+):
+    """Verify per-token-head round-trip: kernel dequant matches reference.
+
+    INT8: Triton truncates on float->int8 store.
+    FP8: hardware cast (clamp then cast).
+    """
+    from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
+        triton_reshape_and_cache_flash_per_token_head_quant,
+    )
+
+    torch.set_default_device("cuda")
+    set_random_seed(42)
+
+    num_blocks = (num_tokens + block_size - 1) // block_size + 2
+
+    key = torch.randn(num_tokens, num_heads, head_size, dtype=torch.bfloat16) * 0.5
+    value = torch.randn(num_tokens, num_heads, head_size, dtype=torch.bfloat16) * 0.5
+
+    key_cache = torch.zeros(
+        num_blocks, block_size, num_heads, head_size, dtype=qcfg.cache_dtype
+    )
+    value_cache = torch.zeros(
+        num_blocks, block_size, num_heads, head_size, dtype=qcfg.cache_dtype
+    )
+    k_scale_cache = torch.ones(num_blocks, block_size, num_heads, dtype=torch.float32)
+    v_scale_cache = torch.ones(num_blocks, block_size, num_heads, dtype=torch.float32)
+
+    slot_mapping = torch.arange(num_tokens, dtype=torch.long)
+
+    triton_reshape_and_cache_flash_per_token_head_quant(
+        key,
+        value,
+        key_cache,
+        value_cache,
+        k_scale_cache,
+        v_scale_cache,
+        slot_mapping,
+    )
+
+    for i in range(num_tokens):
+        blk = i // block_size
+        off = i % block_size
+
+        for label, data, cache, sc in [
+            ("key", key, key_cache, k_scale_cache),
+            ("val", value, value_cache, v_scale_cache),
+        ]:
+            for h in range(num_heads):
+                orig = data[i, h].float()  # [head_size]
+
+                actual_q = cache[blk, off, h]
+                actual_sc = sc[blk, off, h]
+                actual_deq = actual_q.float() * actual_sc
+
+                # Round-trip: dequantized should be close to original
+                torch.testing.assert_close(
+                    actual_deq,
+                    orig,
+                    atol=0.1,
+                    rtol=0.1,
+                )
+
+
+# ===========================================================================
+# 4. Negative slot mapping (padding tokens should be skipped)
+# ===========================================================================
+@torch.inference_mode()
+def test_per_token_head_negative_slot_skipped(qcfg: QuantConfig):
+    """Tokens with slot_mapping=-1 should leave the cache unchanged."""
+    from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
+        triton_reshape_and_cache_flash_per_token_head_quant,
+    )
+
+    torch.set_default_device("cuda")
+    num_tokens = 4
+    num_heads = 2
+    head_size = 64
+    block_size = 16
+    num_blocks = 2
+
+    key = torch.randn(num_tokens, num_heads, head_size, dtype=torch.bfloat16)
+    value = torch.randn(num_tokens, num_heads, head_size, dtype=torch.bfloat16)
+
+    key_cache = torch.zeros(
+        num_blocks, block_size, num_heads, head_size, dtype=qcfg.cache_dtype
+    )
+    value_cache = torch.zeros(
+        num_blocks, block_size, num_heads, head_size, dtype=qcfg.cache_dtype
+    )
+    k_scale_cache = torch.ones(num_blocks, block_size, num_heads, dtype=torch.float32)
+    v_scale_cache = torch.ones(num_blocks, block_size, num_heads, dtype=torch.float32)
+
+    slot_mapping = torch.tensor([0, -1, 1, -1], dtype=torch.long)
+
+    key_cache_before = key_cache.clone()
+    val_cache_before = value_cache.clone()
+
+    triton_reshape_and_cache_flash_per_token_head_quant(
+        key,
+        value,
+        key_cache,
+        value_cache,
+        k_scale_cache,
+        v_scale_cache,
+        slot_mapping,
+    )
+
+    # Slots 0 and 1 should have been written (tokens 0 and 2)
+    assert not torch.equal(key_cache[0, 0], key_cache_before[0, 0])
+    assert not torch.equal(key_cache[0, 1], key_cache_before[0, 1])
+    assert not torch.equal(value_cache[0, 0], val_cache_before[0, 0])
+
+    # All other slots should be unchanged
+    assert torch.equal(key_cache[0, 2:], key_cache_before[0, 2:])
+    assert torch.equal(key_cache[1], key_cache_before[1])
+    assert torch.equal(value_cache[0, 2:], val_cache_before[0, 2:])
+
+
+# ===========================================================================
+# 5. process_weights_after_loading -- per-token-head early return
+# ===========================================================================
+@pytest.mark.parametrize(
+    "kv_cache_dtype", ["int8_per_token_head", "fp8_per_token_head"]
+)
+def test_process_weights_sets_placeholder_scales(kv_cache_dtype: str):
+    """Per-token-head should set _k_scale=1.0, _v_scale=1.0
+    and delete checkpoint attrs."""
+    from vllm.model_executor.layers.quantization.kv_cache import (
+        BaseKVCacheMethod,
+    )
+
+    layer = MagicMock()
+    layer.kv_cache_dtype = kv_cache_dtype
+    layer.calculate_kv_scales = False
+    layer.k_scale = torch.nn.Parameter(torch.tensor(-1.0), requires_grad=False)
+    layer.v_scale = torch.nn.Parameter(torch.tensor(-1.0), requires_grad=False)
+    layer.q_scale = torch.nn.Parameter(torch.tensor(-1.0), requires_grad=False)
+    layer.prob_scale = torch.nn.Parameter(torch.tensor(-1.0), requires_grad=False)
+    layer._k_scale = torch.tensor(0.0)
+    layer._v_scale = torch.tensor(0.0)
+    layer._k_scale_float = 0.0
+    layer._v_scale_float = 0.0
+
+    method = BaseKVCacheMethod.__new__(BaseKVCacheMethod)
+    method.quant_config = MagicMock()
+    method.process_weights_after_loading(layer)
+
+    assert layer._k_scale_float == 1.0
+    assert layer._v_scale_float == 1.0
+    assert not hasattr(layer, "k_scale")
+    assert not hasattr(layer, "v_scale")
+    assert not hasattr(layer, "q_scale")
+    assert not hasattr(layer, "prob_scale")
+
+
+# ===========================================================================
+# 6. Triton unified_attention -- per-token-head scale cache (INT8 and FP8)
+# ===========================================================================
+@pytest.mark.parametrize(
+    "seq_lens",
+    [
+        [(1, 128)],
+        [(1, 64), (1, 32)],
+    ],
+)
+@pytest.mark.parametrize("num_heads", [(4, 4)])
+@pytest.mark.parametrize("head_size", [128])
+@pytest.mark.parametrize("block_size", [16])
+@torch.inference_mode()
+def test_triton_unified_attention_per_token_head_scale(
+    qcfg: QuantConfig,
+    seq_lens: list[tuple[int, int]],
+    num_heads: tuple[int, int],
+    head_size: int,
+    block_size: int,
+):
+    """End-to-end: quantized KV with per-token-head scale caches."""
+    from vllm.utils.math_utils import next_power_of_2
+    from vllm.v1.attention.ops.triton_unified_attention import unified_attention
+
+    torch.set_default_device("cuda")
+    set_random_seed(0)
+
+    num_seqs = len(seq_lens)
+    query_lens = [s[0] for s in seq_lens]
+    kv_lens = [s[1] for s in seq_lens]
+    num_query_heads, num_kv_heads = num_heads
+    max_query_len = max(query_lens)
+    max_kv_len = max(kv_lens)
+    scale = head_size**-0.5
+    num_blocks = 2048
+
+    query = torch.randn(
+        sum(query_lens), num_query_heads, head_size, dtype=torch.bfloat16
+    )
+
+    key_cache_bf16 = torch.randn(
+        num_blocks, block_size, num_kv_heads, head_size, dtype=torch.bfloat16
+    )
+    value_cache_bf16 = torch.randn_like(key_cache_bf16)
+
+    # Per-token-head quantization: one scale per (block, slot, head)
+    k_absmax = key_cache_bf16.float().abs().amax(dim=-1)  # [..., num_kv_heads]
+    v_absmax = value_cache_bf16.float().abs().amax(dim=-1)
+    k_scale_cache = (k_absmax / qcfg.quant_max).clamp(min=1e-6).to(torch.float32)
+    v_scale_cache = (v_absmax / qcfg.quant_max).clamp(min=1e-6).to(torch.float32)
+
+    scaled_k = key_cache_bf16.float() / k_scale_cache[:, :, :, None]
+    scaled_v = value_cache_bf16.float() / v_scale_cache[:, :, :, None]
+    if qcfg.uses_trunc:
+        key_cache_q = (
+            scaled_k.round().clamp(qcfg.quant_min, qcfg.quant_max).to(qcfg.cache_dtype)
+        )
+        value_cache_q = (
+            scaled_v.round().clamp(qcfg.quant_min, qcfg.quant_max).to(qcfg.cache_dtype)
+        )
+    else:
+        key_cache_q = scaled_k.clamp(qcfg.quant_min, qcfg.quant_max).to(
+            qcfg.cache_dtype
+        )
+        value_cache_q = scaled_v.clamp(qcfg.quant_min, qcfg.quant_max).to(
+            qcfg.cache_dtype
+        )
+
+    # Dequantized reference
+    key_cache_deq = key_cache_q.float() * k_scale_cache[:, :, :, None]
+    value_cache_deq = value_cache_q.float() * v_scale_cache[:, :, :, None]
+
+    cu_query_lens = torch.tensor([0] + query_lens, dtype=torch.int32).cumsum(
+        dim=0, dtype=torch.int32
+    )
+    kv_lens_t = torch.tensor(kv_lens, dtype=torch.int32)
+
+    max_num_blocks_per_seq = (max_kv_len + block_size - 1) // block_size
+    block_tables = torch.randint(
+        0, num_blocks, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
+    )
+
+    head_size_padded = next_power_of_2(head_size)
+    seq_threshold_3D = 0
+    num_par_softmax_segments = 16
+    softmax_segm_output = torch.empty(
+        (seq_threshold_3D, num_query_heads, num_par_softmax_segments, head_size_padded),
+        dtype=torch.float32,
+    )
+    softmax_segm_max = torch.empty(
+        (seq_threshold_3D, num_query_heads, num_par_softmax_segments),
+        dtype=torch.float32,
+    )
+    softmax_segm_expsum = torch.empty(
+        (seq_threshold_3D, num_query_heads, num_par_softmax_segments),
+        dtype=torch.float32,
+    )
+
+    output_q = torch.empty_like(query)
+    unified_attention(
+        q=query,
+        k=key_cache_q,
+        v=value_cache_q,
+        out=output_q,
+        cu_seqlens_q=cu_query_lens,
+        seqused_k=kv_lens_t,
+        max_seqlen_q=max_query_len,
+        max_seqlen_k=max_kv_len,
+        softmax_scale=scale,
+        causal=True,
+        window_size=(-1, -1),
+        block_table=block_tables,
+        softcap=0,
+        q_descale=None,
+        k_descale=None,
+        v_descale=None,
+        seq_threshold_3D=seq_threshold_3D,
+        num_par_softmax_segments=num_par_softmax_segments,
+        softmax_segm_output=softmax_segm_output,
+        softmax_segm_max=softmax_segm_max,
+        softmax_segm_expsum=softmax_segm_expsum,
+        kv_quant_mode=qcfg.kv_quant_mode,
+        k_scale_cache=k_scale_cache,
+        v_scale_cache=v_scale_cache,
+    )
+
+    output_ref = torch.empty_like(query)
+    unified_attention(
+        q=query,
+        k=key_cache_deq.to(torch.bfloat16),
+        v=value_cache_deq.to(torch.bfloat16),
+        out=output_ref,
+        cu_seqlens_q=cu_query_lens,
+        seqused_k=kv_lens_t,
+        max_seqlen_q=max_query_len,
+        max_seqlen_k=max_kv_len,
+        softmax_scale=scale,
+        causal=True,
+        window_size=(-1, -1),
+        block_table=block_tables,
+        softcap=0,
+        q_descale=None,
+        k_descale=None,
+        v_descale=None,
+        seq_threshold_3D=seq_threshold_3D,
+        num_par_softmax_segments=num_par_softmax_segments,
+        softmax_segm_output=softmax_segm_output,
+        softmax_segm_max=softmax_segm_max,
+        softmax_segm_expsum=softmax_segm_expsum,
+    )
+
+    torch.testing.assert_close(output_q, output_ref, atol=5e-2, rtol=5e-2)

vllm/config/cache.py

@@ -8,7 +8,10 @@ from pydantic import Field, SkipValidation, field_validator, model_validator
 
 from vllm.config.utils import config
 from vllm.logger import init_logger
-from vllm.utils.torch_utils import is_quantized_kv_cache
+from vllm.utils.torch_utils import (
+    is_quantized_kv_cache,
+    kv_cache_uses_per_token_head_scales,
+)
 
 logger = init_logger(__name__)
 
@@ -21,6 +24,8 @@ CacheDType = Literal[
     "fp8_e5m2",
     "fp8_inc",
     "fp8_ds_mla",
+    "int8_per_token_head",
+    "fp8_per_token_head",
 ]
 MambaDType = Literal["auto", "float32", "float16"]
 MambaCacheMode = Literal["all", "align", "none"]
@@ -237,12 +242,20 @@ class CacheConfig:
     @field_validator("cache_dtype", mode="after")
     @classmethod
     def _validate_cache_dtype(cls, cache_dtype: CacheDType) -> CacheDType:
-        if is_quantized_kv_cache(cache_dtype):
+        if kv_cache_uses_per_token_head_scales(cache_dtype):
             logger.info(
-                "Using fp8 data type to store kv cache. It reduces the GPU "
+                "Using %s data type to store kv cache. It reduces the GPU "
+                "memory footprint and boosts the performance. "
+                "Dynamic per-token-head scales will be computed at runtime.",
+                str(cache_dtype),
+            )
+        elif is_quantized_kv_cache(cache_dtype):
+            logger.info(
+                "Using %s data type to store kv cache. It reduces the GPU "
                 "memory footprint and boosts the performance. "
                 "Meanwhile, it may cause accuracy drop without a proper "
-                "scaling factor."
+                "scaling factor",
+                str(cache_dtype),
             )
         return cache_dtype
 

vllm/model_executor/layers/attention/attention.py

@@ -38,6 +38,7 @@ from vllm.v1.kv_cache_interface import (
     FullAttentionSpec,
     KVCacheSpec,
     SlidingWindowSpec,
+    get_kv_quant_mode,
 )
 
 if TYPE_CHECKING:
@@ -381,8 +382,10 @@ class Attention(nn.Module, AttentionLayerBase):
 
         # for attn backends supporting query quantization
         self.query_quant = None
-        if self.impl.supports_quant_query_input and self.kv_cache_dtype.startswith(
-            "fp8"
+        if (
+            self.impl.supports_quant_query_input
+            and self.kv_cache_dtype.startswith("fp8")
+            and not self.kv_cache_dtype.endswith("per_token_head")
         ):
             is_per_head = (
                 hasattr(self, "q_scale") and self.q_scale.numel() == self.num_kv_heads
@@ -539,6 +542,7 @@ class Attention(nn.Module, AttentionLayerBase):
         block_size = vllm_config.cache_config.block_size
         # Should not be called for enc-dec or encoder-only attention.
         assert self.attn_type == AttentionType.DECODER
+        quant_mode = get_kv_quant_mode(self.kv_cache_dtype)
         if self.sliding_window is not None:
             assert not vllm_config.model_config.use_mla, (
                 "MLA is not supported for slidingwindow"
@@ -548,6 +552,7 @@ class Attention(nn.Module, AttentionLayerBase):
                 num_kv_heads=self.num_kv_heads,
                 head_size=self.head_size,
                 dtype=self.kv_cache_torch_dtype,
+                kv_quant_mode=quant_mode,
                 sliding_window=self.sliding_window,
             )
         else:
@@ -557,6 +562,7 @@ class Attention(nn.Module, AttentionLayerBase):
                 head_size=self.head_size,
                 head_size_v=self.head_size_v,
                 dtype=self.kv_cache_torch_dtype,
+                kv_quant_mode=quant_mode,
             )
 
 

vllm/model_executor/layers/attention/chunked_local_attention.py

@@ -23,6 +23,7 @@ from vllm.v1.kv_cache_interface import (
     AttentionSpec,
     ChunkedLocalAttentionSpec,
     KVCacheSpec,
+    get_kv_quant_mode,
 )
 
 
@@ -123,5 +124,6 @@ class ChunkedLocalAttention(Attention):
             num_kv_heads=self.num_kv_heads,
             head_size=self.head_size,
             dtype=self.kv_cache_torch_dtype,
+            kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype),
             attention_chunk_size=self.attention_chunk_size,
         )

vllm/model_executor/layers/attention/cross_attention.py

@@ -18,7 +18,11 @@ from vllm.v1.attention.backend import (
     subclass_attention_backend_with_overrides,
 )
 from vllm.v1.attention.selector import get_attn_backend
-from vllm.v1.kv_cache_interface import CrossAttentionSpec, KVCacheSpec
+from vllm.v1.kv_cache_interface import (
+    CrossAttentionSpec,
+    KVCacheSpec,
+    get_kv_quant_mode,
+)
 
 logger = init_logger(__name__)
 
@@ -220,4 +224,5 @@ class CrossAttention(Attention):
             num_kv_heads=self.num_kv_heads,
             head_size=self.head_size,
             dtype=self.kv_cache_torch_dtype,
+            kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype),
         )

vllm/model_executor/layers/attention/static_sink_attention.py

@@ -26,6 +26,7 @@ from vllm.v1.kv_cache_interface import (
     AttentionSpec,
     KVCacheSpec,
     SinkFullAttentionSpec,
+    get_kv_quant_mode,
 )
 
 logger = init_logger(__name__)
@@ -217,6 +218,7 @@ class StaticSinkAttention(Attention, CustomOp):
             head_size_v=self.head_size_v,
             sink_len=self.sink_len,
             dtype=self.kv_cache_torch_dtype,
+            kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype),
         )
 
 

vllm/model_executor/layers/quantization/kv_cache.py

@@ -10,6 +10,7 @@ from vllm.model_executor.layers.quantization.base_config import (
 )
 from vllm.platforms import current_platform
 from vllm.utils.torch_utils import is_quantized_kv_cache
+from vllm.v1.kv_cache_interface import kv_cache_uses_per_token_head_scales
 
 logger = init_logger(__name__)
 
@@ -53,6 +54,20 @@ class BaseKVCacheMethod(QuantizeMethodBase):
             assert not hasattr(layer, "prob_scale")
             return
 
+        # Per-token-head quantized KV cache: scales are computed dynamically
+        # per (token, head) in the kernel at cache-write time.  Checkpoint
+        # scales are never used regardless of calculate_kv_scales.
+        if kv_cache_uses_per_token_head_scales(layer.kv_cache_dtype):
+            layer._k_scale.copy_(1.0)
+            layer._v_scale.copy_(1.0)
+            layer._k_scale_float = 1.0
+            layer._v_scale_float = 1.0
+            del layer.k_scale
+            del layer.v_scale
+            del layer.q_scale
+            del layer.prob_scale
+            return
+
         # If the kv-cache is not quantized, we enforce the k/v_scale to be 1.0
         # regardless whether the kv-scale is available in the checkpoint.
         # No need to process kv scales after loading if we are going to

vllm/platforms/interface.py

@@ -505,6 +505,7 @@ class Platform:
             FullAttentionSpec,
             MambaSpec,
             MLAAttentionSpec,
+            get_kv_quant_mode,
         )
 
         cache_config = vllm_config.cache_config
@@ -516,6 +517,8 @@ class Platform:
         else:
             kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
 
+        kv_quant_mode = get_kv_quant_mode(cache_config.cache_dtype)
+
         # Compute attention page size for 1 token
         if model_config.use_mla:
             attn_page_size_1_token = MLAAttentionSpec(
@@ -523,6 +526,7 @@ class Platform:
                 num_kv_heads=model_config.get_num_kv_heads(parallel_config),
                 head_size=model_config.get_head_size(),
                 dtype=kv_cache_dtype,
+                kv_quant_mode=kv_quant_mode,
             ).page_size_bytes
         else:
             attn_page_size_1_token = FullAttentionSpec(
@@ -530,6 +534,7 @@ class Platform:
                 num_kv_heads=model_config.get_num_kv_heads(parallel_config),
                 head_size=model_config.get_head_size(),
                 dtype=kv_cache_dtype,
+                kv_quant_mode=kv_quant_mode,
             ).page_size_bytes
 
         # Compute mamba page size

vllm/utils/torch_utils.py

@@ -37,6 +37,8 @@ STR_DTYPE_TO_TORCH_DTYPE = {
     "fp8_e4m3": torch.uint8,
     "fp8_e5m2": torch.uint8,
     "int8": torch.int8,
+    "int8_per_token_head": torch.int8,
+    "fp8_per_token_head": torch.uint8,
     "fp8_inc": torch.float8_e4m3fn,
     "fp8_ds_mla": torch.uint8,
 }
@@ -62,7 +64,12 @@ T = TypeVar("T")
 
 
 def is_quantized_kv_cache(kv_cache_dtype: str) -> bool:
-    return kv_cache_dtype.startswith("fp8")
+    return kv_cache_dtype.startswith("fp8") or kv_cache_dtype.endswith("per_token_head")
+
+
+def kv_cache_uses_per_token_head_scales(kv_cache_dtype: str) -> bool:
+    """Return True if *kv_cache_dtype* needs per-token-head scales."""
+    return kv_cache_dtype.endswith("per_token_head")
 
 
 def is_strictly_contiguous(t: torch.Tensor) -> bool:

vllm/v1/attention/backend.py

@@ -17,7 +17,9 @@ if TYPE_CHECKING:
     from vllm.model_executor.layers.quantization.utils.quant_utils import QuantKey
     from vllm.platforms.interface import DeviceCapability
     from vllm.v1.attention.backends.utils import KVCacheLayoutType
-    from vllm.v1.kv_cache_interface import AttentionSpec
+    from vllm.v1.kv_cache_interface import AttentionSpec, KVQuantMode
+
+from vllm.v1.kv_cache_interface import get_kv_quant_mode
 
 
 class AttentionType(str, Enum):
@@ -740,6 +742,13 @@ class AttentionImplBase(ABC, Generic[T]):
 class AttentionImpl(AttentionImplBase[T], Generic[T]):
     """Standard attention implementation with forward method."""
 
+    kv_cache_dtype: str
+
+    @property
+    def kv_quant_mode(self) -> "KVQuantMode":
+        """Return the KV cache quantization mode for this layer."""
+        return get_kv_quant_mode(self.kv_cache_dtype)
+
     @abstractmethod
     def __init__(
         self,

vllm/v1/attention/backends/triton_attn.py

@@ -33,9 +33,14 @@ from vllm.v1.attention.backends.utils import get_kv_cache_layout
 from vllm.v1.attention.ops.triton_prefill_attention import context_attention_fwd
 from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
     triton_reshape_and_cache_flash,
+    triton_reshape_and_cache_flash_per_token_head_quant,
 )
 from vllm.v1.attention.ops.triton_unified_attention import unified_attention
-from vllm.v1.kv_cache_interface import AttentionSpec
+from vllm.v1.kv_cache_interface import (
+    AttentionSpec,
+    get_kv_quant_mode,
+    kv_cache_uses_per_token_head_scales,
+)
 
 logger = init_logger(__name__)
 
@@ -270,6 +275,8 @@ class TritonAttentionBackend(AttentionBackend):
         "fp8",
         "fp8_e4m3",
         "fp8_e5m2",
+        "int8_per_token_head",
+        "fp8_per_token_head",
     ]
 
     @staticmethod
@@ -302,6 +309,18 @@ class TritonAttentionBackend(AttentionBackend):
     ) -> tuple[int, ...]:
         if block_size % 16 != 0:
             raise ValueError("Block size must be a multiple of 16.")
+        if kv_cache_uses_per_token_head_scales(cache_dtype_str):
+            # Pad head_size by sizeof(float32)/sizeof(cache_dtype) so
+            # the per-head scale fits inline.  The backend extracts
+            # data[:head_size] and scale[head_size:] via typed views.
+            from vllm.utils.torch_utils import (
+                STR_DTYPE_TO_TORCH_DTYPE,
+                get_dtype_size,
+            )
+
+            cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_dtype_str]
+            scale_pad = get_dtype_size(torch.float32) // get_dtype_size(cache_dtype)
+            return (num_blocks, 2, block_size, num_kv_heads, head_size + scale_pad)
         return (num_blocks, 2, block_size, num_kv_heads, head_size)
 
     @staticmethod
@@ -365,6 +384,62 @@ class TritonAttentionBackend(AttentionBackend):
 
 
 class TritonAttentionImpl(AttentionImpl):
+    # Per-token-head quant: scale views carved from inline head padding.
+    _k_scale_cache: torch.Tensor | None = None
+    _v_scale_cache: torch.Tensor | None = None
+
+    def _ensure_scale_caches(self, kv_cache: torch.Tensor) -> None:
+        """Extract per-head scale views from the padded head dimension.
+
+        The KV cache shape is ``(num_blocks, 2, block_size, nkv, hs+pad)``
+        where ``pad = sizeof(float32) / sizeof(cache_dtype)``.  The last
+        ``pad`` elements of each head hold one float32 scale.  We create
+        strided float32 views over those bytes.
+
+        Scale shape: ``(num_blocks, block_size, num_kv_heads)``
+        """
+        if self._k_scale_cache is not None:
+            return
+        from vllm.utils.torch_utils import get_dtype_size
+
+        num_blocks, _, block_size, nkv, padded_hs = kv_cache.shape
+        dtype_sz = kv_cache.element_size()
+        scale_pad = get_dtype_size(torch.float32) // dtype_sz  # e.g. 4
+        hs = padded_hs - scale_pad
+
+        raw = kv_cache.untyped_storage()
+        base_f32 = torch.tensor([], dtype=torch.float32, device=kv_cache.device).set_(
+            raw
+        )
+
+        # In the raw bytes, each (block, kv_half, slot, head) occupies
+        # padded_hs * dtype_sz bytes.  The scale float32 sits at byte
+        # offset hs * dtype_sz within that region.
+        kv_half_bytes = block_size * nkv * padded_hs * dtype_sz
+        full_block_f32 = 2 * kv_half_bytes // 4  # stride between blocks
+        slot_f32 = nkv * padded_hs * dtype_sz // 4  # stride between slots
+        head_f32 = padded_hs * dtype_sz // 4  # stride between heads
+        scale_off_f32 = hs * dtype_sz // 4  # offset to scale within head
+
+        # K scales: kv_half=0
+        self._k_scale_cache = torch.as_strided(
+            base_f32,
+            size=(num_blocks, block_size, nkv),
+            stride=(full_block_f32, slot_f32, head_f32),
+            storage_offset=scale_off_f32,
+        )
+        self._k_scale_cache.fill_(1.0)
+
+        # V scales: kv_half=1, offset by kv_half_bytes
+        v_base_f32 = kv_half_bytes // 4
+        self._v_scale_cache = torch.as_strided(
+            base_f32,
+            size=(num_blocks, block_size, nkv),
+            stride=(full_block_f32, slot_f32, head_f32),
+            storage_offset=v_base_f32 + scale_off_f32,
+        )
+        self._v_scale_cache.fill_(1.0)
+
     def fused_output_quant_supported(self, quant_key: QuantKey):
         return quant_key == kFp8StaticTensorSym
 
@@ -418,6 +493,9 @@ class TritonAttentionImpl(AttentionImpl):
         self.use_alibi_sqrt = use_alibi_sqrt
         self.supports_quant_query_input = current_platform.is_cuda()
 
+        self._kv_quant_mode = get_kv_quant_mode(kv_cache_dtype)
+        self._is_per_token_head_quant = self._kv_quant_mode.is_per_token_head
+
     def forward(
         self,
         layer: torch.nn.Module,
@@ -480,15 +558,35 @@ class TritonAttentionImpl(AttentionImpl):
                 layer,
             )
 
-        # For decoder and cross-attention, use KV cache as before
-        key_cache, value_cache = kv_cache.unbind(1)
-        if is_quantized_kv_cache(self.kv_cache_dtype):
-            if key_cache.dtype != self.fp8_dtype:
+        # Per-token-head quantized KV cache: use separate scale caches.
+        if self._is_per_token_head_quant:
+            self._ensure_scale_caches(kv_cache)
+            key_cache, value_cache = kv_cache.unbind(1)
+            if key_cache.dtype == torch.uint8:
                 key_cache = key_cache.view(self.fp8_dtype)
                 value_cache = value_cache.view(self.fp8_dtype)
-            assert layer._q_scale_float == 1.0, (
-                "A non 1.0 q_scale is not currently supported."
+            k_descale = None
+            v_descale = None
+            k_scale_cache = self._k_scale_cache
+            v_scale_cache = self._v_scale_cache
+        # FP8 per-tensor / auto path (original flow).
+        else:
+            key_cache, value_cache = kv_cache.unbind(1)
+            if is_quantized_kv_cache(self.kv_cache_dtype):
+                if key_cache.dtype != self.fp8_dtype:
+                    key_cache = key_cache.view(self.fp8_dtype)
+                    value_cache = value_cache.view(self.fp8_dtype)
+                assert layer._q_scale_float == 1.0, (
+                    "A non 1.0 q_scale is not currently supported."
+                )
+            descale_shape = (
+                attn_metadata.query_start_loc.shape[0] - 1,
+                key_cache.shape[2],
             )
+            k_descale = layer._k_scale.expand(descale_shape)
+            v_descale = layer._v_scale.expand(descale_shape)
+            k_scale_cache = None
+            v_scale_cache = None
 
         cu_seqlens_q = attn_metadata.query_start_loc
         seqused_k = attn_metadata.seq_lens
@@ -502,7 +600,6 @@ class TritonAttentionImpl(AttentionImpl):
         softmax_segm_max = attn_metadata.softmax_segm_max
         softmax_segm_expsum = attn_metadata.softmax_segm_expsum
 
-        descale_shape = (cu_seqlens_q.shape[0] - 1, key_cache.shape[2])
         mm_prefix_range_tensor = attn_metadata.mm_prefix_range_tensor
 
         unified_attention(
@@ -522,8 +619,8 @@ class TritonAttentionImpl(AttentionImpl):
             block_table=block_table,
             softcap=self.logits_soft_cap,
             q_descale=None,  # Not supported
-            k_descale=layer._k_scale.expand(descale_shape),
-            v_descale=layer._v_scale.expand(descale_shape),
+            k_descale=k_descale,
+            v_descale=v_descale,
             seq_threshold_3D=seq_threshold_3D,
             num_par_softmax_segments=num_par_softmax_segments,
             softmax_segm_output=softmax_segm_output,
@@ -532,6 +629,9 @@ class TritonAttentionImpl(AttentionImpl):
             sinks=self.sinks,
             output_scale=output_scale,
             mm_prefix_range=mm_prefix_range_tensor,
+            kv_quant_mode=self._kv_quant_mode,
+            k_scale_cache=k_scale_cache,
+            v_scale_cache=v_scale_cache,
         )
 
         return output
@@ -555,10 +655,10 @@ class TritonAttentionImpl(AttentionImpl):
             attn_metadata: Encoder attention metadata
             layer: The attention layer
         """
-        # For encoder attention, process FP8 quantization if needed
+        # Quantized KV cache is not supported for encoder attention.
         if is_quantized_kv_cache(self.kv_cache_dtype):
             raise NotImplementedError(
-                "quantization is not supported for encoder attention"
+                "quantized KV cache is not supported for encoder attention"
             )
 
         # Use encoder-specific metadata for sequence information
@@ -594,16 +694,28 @@ class TritonAttentionImpl(AttentionImpl):
             # For encoder attention,
             # we use direct Q, K, V tensors without caching
             return
-        # For decoder and cross-attention, use KV cache as before
-        key_cache, value_cache = kv_cache.unbind(1)
-
         # Reshape the input keys and values and store them in the cache.
+        if self._is_per_token_head_quant:
+            self._ensure_scale_caches(kv_cache)
+            key_cache, value_cache = kv_cache.unbind(1)
+            if key_cache.dtype == torch.uint8:
+                key_cache = key_cache.view(self.fp8_dtype)
+                value_cache = value_cache.view(self.fp8_dtype)
+            triton_reshape_and_cache_flash_per_token_head_quant(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                self._k_scale_cache,
+                self._v_scale_cache,
+                slot_mapping,
+            )
+            return
+        # For decoder and cross-attention, use KV cache as before.
+        key_cache, value_cache = kv_cache.unbind(1)
         if is_quantized_kv_cache(self.kv_cache_dtype):
             key_cache = key_cache.view(self.fp8_dtype)
             value_cache = value_cache.view(self.fp8_dtype)
-            # triton kernel does not support uint8 kv_cache
-            #  (because some explicit casts (e.g. float8_e4m3fnuz)
-            #   are not supported)
         triton_reshape_and_cache_flash(
             key,
             value,
@@ -616,6 +728,8 @@ class TritonAttentionImpl(AttentionImpl):
         )
 
     def fused_rope_kvcache_supported(self):
+        if self._is_per_token_head_quant:
+            return False
         return rocm_aiter_ops.is_enabled()
 
     def do_rope_and_kv_cache_update(

vllm/v1/attention/ops/triton_reshape_and_cache_flash.py

@@ -3,10 +3,16 @@
 
 import torch
 
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    FP8_DTYPE,
+    get_fp8_min_max,
+)
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
 from vllm.utils.torch_utils import is_quantized_kv_cache
 
+FP8_MIN, FP8_MAX = get_fp8_min_max()
+
 
 @triton.jit
 def reshape_and_cache_kernel_flash(
@@ -118,6 +124,198 @@ def reshape_and_cache_kernel_flash(
     return
 
 
+# ---------------------------------------------------------------------------
+# Per-token-head dynamic quantization kernel
+# Grid: (num_tokens, NUM_KV_HEADS)
+# Each program handles one (token, head) pair:
+#   1. Loads K (or V) for that single head
+#   2. Computes absmax across head_size → scale = absmax / QUANT_MAX
+#   3. Quantizes and stores the data + per-head scale
+#
+# Parametrised by QUANT_MAX / QUANT_MIN so the same code path works
+# for int8 (±127/128), fp8_e4m3 (±448), and other formats.
+# ---------------------------------------------------------------------------
+@triton.jit
+def _reshape_cache_per_token_head(
+    key_ptr,  # [num_tokens, num_kv_heads, head_size]
+    value_ptr,  # [num_tokens, num_kv_heads, head_size_v]
+    key_cache_ptr,  # [num_blocks, block_size, num_kv_heads, head_size]
+    value_cache_ptr,  # [num_blocks, block_size, num_kv_heads, head_size_v]
+    k_scale_cache_ptr,  # [num_blocks, block_size, num_kv_heads] float32
+    v_scale_cache_ptr,  # [num_blocks, block_size, num_kv_heads] float32
+    slot_mapping_ptr,  # [num_tokens]
+    stride_key_tok: tl.int64,
+    stride_key_head: tl.int64,
+    stride_val_tok: tl.int64,
+    stride_val_head: tl.int64,
+    stride_kc_blk: tl.int64,  # key_cache stride over blocks
+    stride_kc_slot: tl.int64,  # key_cache stride over slots
+    stride_kc_head: tl.int64,  # key_cache stride over heads
+    stride_vc_blk: tl.int64,
+    stride_vc_slot: tl.int64,
+    stride_vc_head: tl.int64,
+    stride_ks_blk: tl.int64,  # k_scale_cache stride[0] (blocks)
+    stride_ks_slot: tl.int64,  # k_scale_cache stride[1] (slots)
+    stride_ks_head: tl.int64,  # k_scale_cache stride[2] (heads)
+    stride_vs_blk: tl.int64,  # v_scale_cache stride[0] (blocks)
+    stride_vs_slot: tl.int64,  # v_scale_cache stride[1] (slots)
+    stride_vs_head: tl.int64,  # v_scale_cache stride[2] (heads)
+    block_size: tl.constexpr,
+    head_size: tl.constexpr,
+    head_size_v: tl.constexpr,
+    HEAD_SIZE_PADDED: tl.constexpr,  # next_power_of_2(max(head_size, head_size_v))
+    QUANT_MAX: tl.constexpr = 127.0,
+    QUANT_MIN: tl.constexpr = -128.0,
+):
+    tok = tl.program_id(0)
+    head = tl.program_id(1)
+
+    slot = tl.load(slot_mapping_ptr + tok).to(tl.int64)
+    if slot < 0:
+        return
+
+    blk = slot // block_size
+    slot_in_blk = slot % block_size
+
+    dim_offs = tl.arange(0, HEAD_SIZE_PADDED)
+
+    # ---- Key: load one head → absmax → quantize → store -------------------
+    k_mask = dim_offs < head_size
+    k_h = tl.load(
+        key_ptr + tok * stride_key_tok + head * stride_key_head + dim_offs,
+        mask=k_mask,
+        other=0.0,
+    ).to(tl.float32)
+
+    k_scale = tl.maximum(tl.max(tl.abs(k_h)) / QUANT_MAX, 1e-6)
+    tl.store(
+        k_scale_cache_ptr
+        + blk * stride_ks_blk
+        + slot_in_blk * stride_ks_slot
+        + head * stride_ks_head,
+        k_scale,
+    )
+
+    k_q = tl.clamp(k_h * (1.0 / k_scale), QUANT_MIN, QUANT_MAX)
+    tl.store(
+        key_cache_ptr
+        + blk * stride_kc_blk
+        + slot_in_blk * stride_kc_slot
+        + head * stride_kc_head
+        + dim_offs,
+        k_q,
+        mask=k_mask,
+    )
+
+    # ---- Value: same per-head approach ------------------------------------
+    v_mask = dim_offs < head_size_v
+    v_h = tl.load(
+        value_ptr + tok * stride_val_tok + head * stride_val_head + dim_offs,
+        mask=v_mask,
+        other=0.0,
+    ).to(tl.float32)
+
+    v_scale = tl.maximum(tl.max(tl.abs(v_h)) / QUANT_MAX, 1e-6)
+    tl.store(
+        v_scale_cache_ptr
+        + blk * stride_vs_blk
+        + slot_in_blk * stride_vs_slot
+        + head * stride_vs_head,
+        v_scale,
+    )
+
+    v_q = tl.clamp(v_h * (1.0 / v_scale), QUANT_MIN, QUANT_MAX)
+    tl.store(
+        value_cache_ptr
+        + blk * stride_vc_blk
+        + slot_in_blk * stride_vc_slot
+        + head * stride_vc_head
+        + dim_offs,
+        v_q,
+        mask=v_mask,
+    )
+
+
+# Mapping from cache torch dtype to (QUANT_MAX, QUANT_MIN) for the
+# per-token-head quantization kernel.
+_PER_TOKEN_HEAD_QUANT_PARAMS: dict[torch.dtype, tuple[float, float]] = {
+    torch.int8: (127.0, -128.0),
+    FP8_DTYPE: (FP8_MAX, FP8_MIN),
+}
+
+
+def triton_reshape_and_cache_flash_per_token_head_quant(
+    key: torch.Tensor,  # [num_tokens, num_kv_heads, head_size]
+    value: torch.Tensor,  # [num_tokens, num_kv_heads, head_size_v]
+    key_cache: torch.Tensor,  # [num_blocks, block_size, num_kv_heads, head_size]
+    value_cache: torch.Tensor,  # [num_blocks, block_size, num_kv_heads, head_size_v]
+    k_scale_cache: torch.Tensor,  # [num_blocks, block_size, num_kv_heads] float32
+    v_scale_cache: torch.Tensor,  # [num_blocks, block_size, num_kv_heads] float32
+    slot_mapping: torch.Tensor,  # [num_tokens]
+):
+    """Quantize key/value per (token, head) and write to paged cache.
+
+    Computes one scale = absmax / QUANT_MAX per (token, head), stores
+    quantized data in key_cache/value_cache, and stores the float32
+    scale in k_scale_cache/v_scale_cache.
+
+    The quantization range (QUANT_MAX, QUANT_MIN) is derived from the
+    cache tensor dtype so the same code path works for int8 and fp8.
+    """
+    cache_dtype = key_cache.dtype
+    quant_params = _PER_TOKEN_HEAD_QUANT_PARAMS.get(cache_dtype)
+    if quant_params is None:
+        raise ValueError(
+            f"Per-token-head quantization not supported for cache dtype "
+            f"{cache_dtype}.  Supported: {list(_PER_TOKEN_HEAD_QUANT_PARAMS)}"
+        )
+    quant_max, quant_min = quant_params
+
+    num_tokens, num_kv_heads, head_size = key.shape
+    head_size_v = value.shape[2]
+    head_size_padded = triton.next_power_of_2(max(head_size, head_size_v))
+
+    block_size = key_cache.shape[1]
+
+    if current_platform.is_rocm() or current_platform.is_xpu():
+        num_warps = 4
+    else:
+        num_warps = min(16, max(1, head_size_padded // 32))
+
+    _reshape_cache_per_token_head[(num_tokens, num_kv_heads)](
+        key_ptr=key,
+        value_ptr=value,
+        key_cache_ptr=key_cache,
+        value_cache_ptr=value_cache,
+        k_scale_cache_ptr=k_scale_cache,
+        v_scale_cache_ptr=v_scale_cache,
+        slot_mapping_ptr=slot_mapping,
+        stride_key_tok=key.stride(0),
+        stride_key_head=key.stride(1),
+        stride_val_tok=value.stride(0),
+        stride_val_head=value.stride(1),
+        stride_kc_blk=key_cache.stride(0),
+        stride_kc_slot=key_cache.stride(1),
+        stride_kc_head=key_cache.stride(2),
+        stride_vc_blk=value_cache.stride(0),
+        stride_vc_slot=value_cache.stride(1),
+        stride_vc_head=value_cache.stride(2),
+        stride_ks_blk=k_scale_cache.stride(0),
+        stride_ks_slot=k_scale_cache.stride(1),
+        stride_ks_head=k_scale_cache.stride(2),
+        stride_vs_blk=v_scale_cache.stride(0),
+        stride_vs_slot=v_scale_cache.stride(1),
+        stride_vs_head=v_scale_cache.stride(2),
+        block_size=block_size,
+        head_size=head_size,
+        head_size_v=head_size_v,
+        HEAD_SIZE_PADDED=head_size_padded,
+        QUANT_MAX=quant_max,
+        QUANT_MIN=quant_min,
+        num_warps=num_warps,
+    )
+
+
 def triton_reshape_and_cache_flash(
     key: torch.Tensor,  # [num_tokens, num_heads, head_size]
     value: torch.Tensor,  # [num_tokens, num_heads, head_size]
@@ -224,7 +422,6 @@ def triton_reshape_and_cache_flash(
         block_size=block_size,
         x=x,
         USE_HEAD_MAJOR_LAYOUT=use_head_major_layout,
-        # FP8 flags
         FP8_KV_CACHE=FP8_KV_CACHE,
         # autotune parameters
         TILE_SIZE=TILE_SIZE,

vllm/v1/attention/ops/triton_unified_attention.py

@@ -13,6 +13,7 @@ import vllm.envs as envs
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
+from vllm.v1.kv_cache_interface import KVQuantMode
 
 logger = init_logger(__name__)
 is_batch_invariant = envs.VLLM_BATCH_INVARIANT
@@ -32,6 +33,63 @@ def apply_softcap(S, x):
     return x * (p1 - p2) / (p1 + p2)
 
 
+@triton.jit
+def _prepare_kv_tile(
+    data,
+    Q,
+    tensor_scale,
+    scale_cache_ptr,
+    physical_block_idx,
+    seq_offset,
+    kv_head_idx,
+    stride_s_blk,
+    stride_s_slot,
+    stride_s_head,
+    tile_mask,
+    BLOCK_SIZE: tl.constexpr,
+    KV_QUANT_MODE: tl.constexpr,
+):
+    """Prepare a loaded KV tile for attention computation.
+
+    Casts the raw KV data to Q's dtype and loads per-token-head scales
+    when applicable:
+
+    - ``KV_QUANT_MODE == 0``: cast only (no-op for bf16/fp16).
+    - ``KV_QUANT_MODE == 1`` (FP8 per-tensor): dequantize inline
+      using the tensor-wide scale.
+    - ``KV_QUANT_MODE >= 2`` (per-token-head int8/fp8): cast to Q's
+      dtype and return per-head scales separately — the caller applies
+      them after the dot product for better numerical efficiency.
+
+    Returns ``(data, token_head_scales)``.  *token_head_scales* is only
+    meaningful when ``KV_QUANT_MODE >= 2``; callers gate its use on
+    the same constexpr so the compiler eliminates dead code.
+    """
+    # KV_QUANT_MODE values: 0=none, 1=fp8 per-tensor,
+    #                       2=int8 per-token-head, 3=fp8 per-token-head
+
+    # Placeholder scales (float32) — never read when KV_QUANT_MODE < 2.
+    unused_scales = tile_mask.to(tl.float32)
+
+    if KV_QUANT_MODE == 1:  # FP8 per-tensor
+        if Q.dtype.is_fp8():
+            return data.to(Q.dtype), unused_scales
+        return (data.to(tl.float32) * tl.load(tensor_scale)).to(Q.dtype), unused_scales
+    if KV_QUANT_MODE >= 2:  # per-token-head (int8 or fp8)
+        scale_idx = (
+            physical_block_idx * stride_s_blk
+            + (seq_offset % BLOCK_SIZE) * stride_s_slot
+            + kv_head_idx * stride_s_head
+        )
+        token_head_scales = tl.load(
+            scale_cache_ptr + scale_idx, mask=tile_mask, other=1.0
+        )
+        return data.to(Q.dtype), token_head_scales
+    # .to(Q.dtype) is a no-op when data is already Q's type (bf16/fp16),
+    # but required so Triton sees consistent return types across branches.
+    return data.to(Q.dtype), unused_scales
+
+
 @triton.jit
 def find_seq_idx(
     query_start_len_ptr,
@@ -105,8 +163,20 @@ def kernel_unified_attention_2d(
     num_seqs: tl.int32,
     BLOCK_M: tl.constexpr,  # int
     USE_FP8: tl.constexpr,  # bool
+    # KV cache quantization: 0=none, 1=fp8, 2=per-token-head
+    KV_QUANT_MODE: tl.constexpr = 0,
     FP8_MIN: tl.constexpr = float8_info.min,
     FP8_MAX: tl.constexpr = float8_info.max,
+    # Per-token-head scale caches (KV_QUANT_MODE >= 2)
+    # Shape: [num_blocks, block_size, num_kv_heads]
+    k_scale_cache_ptr=None,
+    v_scale_cache_ptr=None,
+    stride_ks_blk=0,
+    stride_ks_slot=0,
+    stride_ks_head=0,
+    stride_vs_blk=0,
+    stride_vs_slot=0,
+    stride_vs_head=0,
 ):
     q_block_global_idx = tl.program_id(0)
     kv_head_idx = tl.program_id(1)
@@ -258,14 +328,21 @@ def kernel_unified_attention_2d(
             mask=dim_mask[:, None] & tile_mask[None, :],
             other=0.0,
         )
-
-        if K_load.dtype.is_fp8():
-            if Q.dtype.is_fp8():
-                K = K_load
-            else:
-                K = (K_load.to(tl.float32) * tl.load(k_scale)).to(Q.dtype)
-        else:
-            K = K_load
+        K, k_token_head_scales = _prepare_kv_tile(
+            K_load,
+            Q,
+            k_scale,
+            k_scale_cache_ptr,
+            physical_block_idx,
+            seq_offset,
+            kv_head_idx,
+            stride_ks_blk,
+            stride_ks_slot,
+            stride_ks_head,
+            tile_mask,
+            BLOCK_SIZE,
+            KV_QUANT_MODE,
+        )
 
         # V : (TILE_SIZE, HEAD_SIZE)
         V_load = tl.load(
@@ -273,14 +350,21 @@ def kernel_unified_attention_2d(
             mask=dim_mask[None, :] & tile_mask[:, None],
             other=0.0,
         )
-
-        if V_load.dtype.is_fp8():
-            if Q.dtype.is_fp8():
-                V = V_load
-            else:
-                V = (V_load.to(tl.float32) * tl.load(v_scale)).to(Q.dtype)
-        else:
-            V = V_load
+        V, v_token_head_scales = _prepare_kv_tile(
+            V_load,
+            Q,
+            v_scale,
+            v_scale_cache_ptr,
+            physical_block_idx,
+            seq_offset,
+            kv_head_idx,
+            stride_vs_blk,
+            stride_vs_slot,
+            stride_vs_head,
+            tile_mask,
+            BLOCK_SIZE,
+            KV_QUANT_MODE,
+        )
 
         # Compute attention mask: causal by default (key <= query)
         query_abs_pos = context_len + query_pos[:, None]
@@ -318,7 +402,12 @@ def kernel_unified_attention_2d(
         # S : (BLOCK_M, TILE_SIZE)
         S = tl.zeros(shape=(BLOCK_M, TILE_SIZE), dtype=tl.float32)
 
-        S += scale * tl.dot(Q, K)
+        # Per-token-head quant: fuse softmax_scale with per-head k_scale
+        # to avoid a separate BLOCK_M × TILE_SIZE multiply on S.
+        if KV_QUANT_MODE >= 2:
+            S += tl.dot(Q, K) * (scale * k_token_head_scales[None, :])
+        else:
+            S += scale * tl.dot(Q, K)
 
         if USE_SOFTCAP:
             S = apply_softcap(S, softcap)
@@ -382,7 +471,12 @@ def kernel_unified_attention_2d(
             )
 
         # acc : (BLOCK_M, HEAD_SIZE_PADDED)
-        acc += tl.dot(P.to(V.dtype), V)
+        # Per-token-head quant: apply v_scale to P instead of V.
+        if KV_QUANT_MODE >= 2:
+            P_v = (P * v_token_head_scales[None, :]).to(V.dtype)
+            acc += tl.dot(P_v, V)
+        else:
+            acc += tl.dot(P.to(V.dtype), V)
 
     # epilogue
     acc = acc / L[:, None]
@@ -453,6 +547,18 @@ def kernel_unified_attention_3d(
     USE_MM_PREFIX: tl.constexpr,  # bool
     MAX_MM_RANGES: tl.constexpr,  # int
     mm_prefix_range_ptr,  # [num_seqs] - prefix length for each sequence
+    # KV cache quantization: 0=none, 1=fp8, 2=per-token-head
+    KV_QUANT_MODE: tl.constexpr = 0,
+    # Per-token-head scale caches (KV_QUANT_MODE >= 2)
+    # Shape: [num_blocks, block_size, num_kv_heads]
+    k_scale_cache_ptr=None,
+    v_scale_cache_ptr=None,
+    stride_ks_blk=0,
+    stride_ks_slot=0,
+    stride_ks_head=0,
+    stride_vs_blk=0,
+    stride_vs_slot=0,
+    stride_vs_head=0,
 ):
     q_block_global_idx = tl.program_id(0)
     kv_head_idx = tl.program_id(1)
@@ -613,14 +719,21 @@ def kernel_unified_attention_3d(
             mask=dim_mask[:, None] & tile_mask[None, :],
             other=0.0,
         )
-
-        if K_load.dtype.is_fp8():
-            if Q.dtype.is_fp8():
-                K = K_load
-            else:
-                K = (K_load.to(tl.float32) * tl.load(k_scale)).to(Q.dtype)
-        else:
-            K = K_load
+        K, k_token_head_scales = _prepare_kv_tile(
+            K_load,
+            Q,
+            k_scale,
+            k_scale_cache_ptr,
+            physical_block_idx,
+            seq_offset,
+            kv_head_idx,
+            stride_ks_blk,
+            stride_ks_slot,
+            stride_ks_head,
+            tile_mask,
+            BLOCK_SIZE,
+            KV_QUANT_MODE,
+        )
 
         # V : (TILE_SIZE, HEAD_SIZE)
         V_load = tl.load(
@@ -628,14 +741,21 @@ def kernel_unified_attention_3d(
             mask=dim_mask[None, :] & tile_mask[:, None],
             other=0.0,
         )
-
-        if V_load.dtype.is_fp8():
-            if Q.dtype.is_fp8():
-                V = V_load
-            else:
-                V = (V_load.to(tl.float32) * tl.load(v_scale)).to(Q.dtype)
-        else:
-            V = V_load
+        V, v_token_head_scales = _prepare_kv_tile(
+            V_load,
+            Q,
+            v_scale,
+            v_scale_cache_ptr,
+            physical_block_idx,
+            seq_offset,
+            kv_head_idx,
+            stride_vs_blk,
+            stride_vs_slot,
+            stride_vs_head,
+            tile_mask,
+            BLOCK_SIZE,
+            KV_QUANT_MODE,
+        )
 
         # Compute attention mask: causal by default (key <= query)
         query_abs_pos = context_len + query_pos[:, None]
@@ -672,7 +792,13 @@ def kernel_unified_attention_3d(
 
         # S : (BLOCK_M, TILE_SIZE)
         S = tl.zeros(shape=(BLOCK_M, TILE_SIZE), dtype=tl.float32)
-        S += scale * tl.dot(Q, K)
+
+        # Per-token-head quant: fuse softmax_scale with per-head k_scale
+        # to avoid a separate BLOCK_M × TILE_SIZE multiply on S.
+        if KV_QUANT_MODE >= 2:
+            S += tl.dot(Q, K) * (scale * k_token_head_scales[None, :])
+        else:
+            S += scale * tl.dot(Q, K)
 
         if USE_SOFTCAP:
             S = apply_softcap(S, softcap)
@@ -736,7 +862,12 @@ def kernel_unified_attention_3d(
             )
 
         # acc : (BLOCK_M, HEAD_SIZE_PADDED)
-        acc += tl.dot(P.to(V.dtype), V)
+        # Per-token-head quant: apply v_scale to P instead of V.
+        if KV_QUANT_MODE >= 2:
+            P_v = (P * v_token_head_scales[None, :]).to(V.dtype)
+            acc += tl.dot(P_v, V)
+        else:
+            acc += tl.dot(P.to(V.dtype), V)
 
     segm_output_offset = (
         query_offset_0[:, None].to(tl.int64)
@@ -911,6 +1042,10 @@ def unified_attention(
     # Optional tensor for prefix lengths (PrefixLM support)
     mm_prefix_range=None,
     use_alibi_sqrt=False,
+    # KV cache quantization mode and per-token-head scale caches.
+    kv_quant_mode: KVQuantMode = KVQuantMode.NONE,
+    k_scale_cache=None,  # [num_blocks, block_size, num_kv_heads] float32
+    v_scale_cache=None,  # [num_blocks, block_size, num_kv_heads] float32
 ):
     assert causal, "Only causal attention is supported"
     assert q_descale is None, "Q scales not supported"
@@ -1040,6 +1175,15 @@ def unified_attention(
             num_seqs=num_seqs,
             BLOCK_M=BLOCK_M,
             USE_FP8=output_scale is not None,
+            KV_QUANT_MODE=kv_quant_mode,
+            k_scale_cache_ptr=k_scale_cache,
+            v_scale_cache_ptr=v_scale_cache,
+            stride_ks_blk=k_scale_cache.stride(0) if k_scale_cache is not None else 0,
+            stride_ks_slot=k_scale_cache.stride(1) if k_scale_cache is not None else 0,
+            stride_ks_head=k_scale_cache.stride(2) if k_scale_cache is not None else 0,
+            stride_vs_blk=v_scale_cache.stride(0) if v_scale_cache is not None else 0,
+            stride_vs_slot=v_scale_cache.stride(1) if v_scale_cache is not None else 0,
+            stride_vs_head=v_scale_cache.stride(2) if v_scale_cache is not None else 0,
         )
     else:
         kernel_unified_attention_3d[
@@ -1092,6 +1236,15 @@ def unified_attention(
             num_seqs=num_seqs,
             BLOCK_M=BLOCK_M,
             NUM_SEGMENTS_PER_SEQ=num_par_softmax_segments,
+            KV_QUANT_MODE=kv_quant_mode,
+            k_scale_cache_ptr=k_scale_cache,
+            v_scale_cache_ptr=v_scale_cache,
+            stride_ks_blk=k_scale_cache.stride(0) if k_scale_cache is not None else 0,
+            stride_ks_slot=k_scale_cache.stride(1) if k_scale_cache is not None else 0,
+            stride_ks_head=k_scale_cache.stride(2) if k_scale_cache is not None else 0,
+            stride_vs_blk=v_scale_cache.stride(0) if v_scale_cache is not None else 0,
+            stride_vs_slot=v_scale_cache.stride(1) if v_scale_cache is not None else 0,
+            stride_vs_head=v_scale_cache.stride(2) if v_scale_cache is not None else 0,
         )
         reduce_segments[(q.shape[0], num_query_heads)](
             output_ptr=out,

vllm/v1/kv_cache_interface.py

@@ -1,21 +1,70 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+from __future__ import annotations
+
 import copy
 from dataclasses import dataclass, fields, replace
+from enum import IntEnum
 from math import prod
+from typing import TYPE_CHECKING
 
 import torch
 from typing_extensions import Self
 
-from vllm.config import VllmConfig
 from vllm.logger import init_logger
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
 from vllm.utils.math_utils import cdiv
 from vllm.utils.torch_utils import get_dtype_size
 
 logger = init_logger(__name__)
 
 
+# ---------------------------------------------------------------------------
+# KV cache quantization mode
+# ---------------------------------------------------------------------------
+
+
+class KVQuantMode(IntEnum):
+    """KV cache quantization mode.
+
+    Used by attention backends and kernels to dispatch quantization logic
+    without string matching on ``kv_cache_dtype``.
+    """
+
+    NONE = 0
+    FP8_PER_TENSOR = 1  # per-tensor scales (current fp8 path)
+    INT8_PER_TOKEN_HEAD = 2  # per-token-head dynamic scales for int8
+    FP8_PER_TOKEN_HEAD = 3  # per-token-head dynamic scales for fp8
+
+    @property
+    def is_per_token_head(self) -> bool:
+        """True for any per-token-head quantization mode."""
+        return self >= 2
+
+
+def get_kv_quant_mode(kv_cache_dtype: str) -> KVQuantMode:
+    """Map a ``kv_cache_dtype`` string to a :class:`KVQuantMode`."""
+    if kv_cache_dtype == "int8_per_token_head":
+        return KVQuantMode.INT8_PER_TOKEN_HEAD
+    if kv_cache_dtype == "fp8_per_token_head":
+        return KVQuantMode.FP8_PER_TOKEN_HEAD
+    if kv_cache_dtype.startswith("fp8"):
+        return KVQuantMode.FP8_PER_TENSOR
+    return KVQuantMode.NONE
+
+
+def is_quantized_kv_cache(kv_cache_dtype: str) -> bool:
+    return get_kv_quant_mode(kv_cache_dtype) != KVQuantMode.NONE
+
+
+def kv_cache_uses_per_token_head_scales(kv_cache_dtype: str) -> bool:
+    """Return True if *kv_cache_dtype* needs per-token-head scales."""
+    return get_kv_quant_mode(kv_cache_dtype).is_per_token_head
+
+
 @dataclass(frozen=True)
 class KVCacheSpec:
     """
@@ -66,11 +115,19 @@ class AttentionSpec(KVCacheSpec):
     num_kv_heads: int
     head_size: int
     dtype: torch.dtype
+    kv_quant_mode: KVQuantMode = KVQuantMode.NONE
     page_size_padded: int | None = None
 
     @property
     def page_size_bytes(self) -> int:
         real_page_size = self.real_page_size_bytes
+        # Per-token-head scales are stored in separate tensors managed
+        # by the attention backend, but the memory is carved from the
+        # raw KV cache allocation so it must be budgeted here.
+        if self.kv_quant_mode.is_per_token_head:
+            real_page_size += (
+                2 * self.block_size * self.num_kv_heads * get_dtype_size(torch.float32)
+            )
         if self.page_size_padded is not None:
             assert self.page_size_padded >= real_page_size
             return self.page_size_padded
@@ -159,6 +216,7 @@ class FullAttentionSpec(AttentionSpec):
             head_size=specs[0].head_size,
             head_size_v=specs[0].head_size_v,
             dtype=specs[0].dtype,
+            kv_quant_mode=specs[0].kv_quant_mode,
             page_size_padded=specs[0].page_size_padded,
             sliding_window=cls.merge_window_sizes(sliding_window),
             attention_chunk_size=cls.merge_window_sizes(attention_chunk_size),
@@ -220,6 +278,7 @@ class MLAAttentionSpec(FullAttentionSpec):
             num_kv_heads=specs[0].num_kv_heads,
             head_size=specs[0].head_size,
             dtype=specs[0].dtype,
+            kv_quant_mode=specs[0].kv_quant_mode,
             page_size_padded=specs[0].page_size_padded,
             cache_dtype_str=cache_dtype_str_set.pop(),
         )
@@ -352,6 +411,7 @@ class SinkFullAttentionSpec(FullAttentionSpec):
             head_size_v=specs[0].head_size_v,
             sink_len=specs[0].sink_len,
             dtype=specs[0].dtype,
+            kv_quant_mode=specs[0].kv_quant_mode,
             page_size_padded=specs[0].page_size_padded,
             sliding_window=cls.merge_window_sizes(sliding_window),
             attention_chunk_size=cls.merge_window_sizes(attention_chunk_size),