Skip to content

[Feature] Triton INT4 / INT2 per-token-head KV cache quantization #40633

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
JartX wants to merge 7 commits into from
Closed

[Feature] Triton INT4 / INT2 per-token-head KV cache quantization #40633

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
d485ea1
feat: INT4 / INT2 per-token-head KV cache quantization
JartX Apr 23, 2026
b3a14c6
refactor(triton_quant_kv): split packed_per_token_head.py
JartX Apr 24, 2026
d67bcf4
Merge branch 'main' into feature/int4-int2-per-token-head-v2
JartX Apr 25, 2026
8e129ec
Merge branch 'main' into feature/int4-int2-per-token-head-v2
JartX Apr 28, 2026
48708c8
Merge branch 'main' into feature/int4-int2-per-token-head-v2
JartX May 4, 2026
47de506
fix: INT8 per-token-head KV cache quantizer rounds to nearest
JartX May 12, 2026
4d33ce1
perf(rdna3): dtype-aware RHT sign caching in _hadamard.py
JartX May 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion docs/design/attention_backends.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -179,7 +179,7 @@ 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`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `int2_per_token_head`, `int4_per_token_head`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
| `TURBOQUANT` | | fp16, bf16 | `turboquant_k8v4`, `turboquant_4bit_nc`, `turboquant_k3v4_nc`, `turboquant_3bit_nc` | 16, 32, 64, 128 | Any | ❌ | ❌ | ❌ | Decoder | Any |

> **†** FlashInfer uses TRTLLM attention on Blackwell (SM100), which supports sinks. Disable via `--attention-config.use_trtllm_attention=0`.
Expand Down
21 changes: 17 additions & 4 deletions tests/models/quantization/test_per_token_kv_cache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,17 +22,30 @@
reason="Per-token-head KV cache requires CUDA or ROCm GPU.",
)
@pytest.mark.parametrize(
"base_model,test_model",
"base_model,test_model,kv_cache_dtype",
[
(
"meta-llama/Meta-Llama-3-8B",
"meta-llama/Meta-Llama-3-8B",
"int2_per_token_head",
),
(
"meta-llama/Meta-Llama-3-8B",
"meta-llama/Meta-Llama-3-8B",
"int4_per_token_head",
),
(
"meta-llama/Llama-3.2-1B-Instruct",
"meta-llama/Llama-3.2-1B-Instruct",
"int8_per_token_head",
),
(
"meta-llama/Llama-3.2-1B-Instruct",
"meta-llama/Llama-3.2-1B-Instruct",
"fp8_per_token_head",
),
],
)
@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"])
Expand Down
Loading