This PR addresses a bug in the Cutlass integration where the
`sparsity_config.ignore` list was not being respected. When only a
subset of modules were configured as Sparse24, the system incorrectly
selected Cutlass for non-sparse modules as well. This update ensures the
correct scheme is selected for non-sparse modules, fixing this behavior.
---
### Changes
- Updated logic to correctly respect `sparsity_config.ignore`.
- Ensured non-sparse modules use the appropriate scheme instead of
defaulting to Cutlass.
---
<details>
<summary>Testing Setup</summary>
The fix has been tested on top of [this
diff](https://github.com/vllm-project/vllm/pull/12097).
#### Steps to Test:
```bash
git checkout -b my-test-branch origin/rahul-bitmask-additions # compressed Cutlass support
git revert --no-edit aa2cd2c # revert Tyler's commit to turn off Cutlass for W16A16
git cherry-pick ca624cddb # this branch
```
#### Additional Patch Required:
```diff
diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
index a54177c1c..f916dd0c9 100644
--- a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
+++ b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
@@ -9,7 +9,7 @@ from compressed_tensors.quantization import (QuantizationArgs,
QuantizationStrategy,
QuantizationType)
from pydantic import BaseModel
-
+from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
UnquantizedLinearMethod)
@@ -27,7 +27,7 @@ from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
should_ignore_layer)
from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
from vllm.platforms import current_platform
-
+logger = init_logger(__name__)
__all__ = ["CompressedTensorsLinearMethod"]
SPARSITY_CONFIG_NAME: Literal["sparsity_config"] = "sparsity_config"
```
Apply using:
```bash
git apply logging-patch.patch
```
</details>
---
<details>
<summary>Models Tested</summary>
- `nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-partial-24`
- `nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-full-sparse24`
-
`nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-partial-24-entire-fp8-compressed`
-
`nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-partial-24-remaining-fp8-compressed`
</details>
---
<details>
<summary>Example Output</summary>
#### Layers 0-5 (Sparse24)
```
Using scheme: CompressedTensors24 for model.layers.0.self_attn.qkv_proj
Using scheme: CompressedTensors24 for model.layers.0.self_attn.o_proj
Using scheme: CompressedTensors24 for model.layers.0.mlp.gate_up_proj
Using scheme: CompressedTensors24 for model.layers.0.mlp.down_proj
...
```
#### Layers 6+ (Non-Sparse, FP8)
```
Using scheme: CompressedTensorsW8A8Fp8 for model.layers.6.self_attn.qkv_proj
Using scheme: CompressedTensorsW8A8Fp8 for model.layers.6.self_attn.o_proj
Using scheme: CompressedTensorsW8A8Fp8 for model.layers.6.mlp.gate_up_proj
Using scheme: CompressedTensorsW8A8Fp8 for model.layers.6.mlp.down_proj
...
```
</details>
**Note:** Assumed all modules in fused layers such as `QKV_proj` and
`Gate_up_proj` follow the same quantization/pruning scheme.
---
For related tasks using the Asana app for GitHub, refer to [[this
link](https://app.asana.com/0/0/1209227810815160)](https://app.asana.com/0/0/1209227810815160).
Signed-off-by: Rahul Tuli <rahul@neuralmagic.com>
Without this PR
---------------
Quantizing models with llm-compressor and a recipe that explicitly lists
names of layers produces a model that is not loadable by vLLM (i.e.
`vllm serve <model>` fails with `raise ValueError(f"Unable to find
matching target for {module} in the ...`).
Example recipe:
```
recipe = """
quantization_stage:
run_type: oneshot
quantization_modifiers:
GPTQModifier:
ignore: ["lm_head"]
config_groups:
group_0:
weights:
num_bits: 4
type: "int"
symmetric: true
strategy: "group"
group_size: 128
targets: [
"model.layers.0.mlp.down_proj",
"model.layers.2.mlp.down_proj",
"model.layers.3.mlp.down_proj",
"model.layers.4.mlp.down_proj",
"model.layers.5.mlp.down_proj",
"model.layers.6.mlp.down_proj",
"model.layers.7.mlp.down_proj",
"model.layers.8.mlp.down_proj",
"model.layers.9.mlp.down_proj",
"model.layers.10.mlp.down_proj",
"model.layers.11.mlp.down_proj",
"model.layers.12.mlp.down_proj",
"model.layers.13.mlp.down_proj",
"model.layers.14.mlp.down_proj",
"model.layers.15.mlp.down_proj",
"model.layers.16.mlp.down_proj",
"model.layers.17.mlp.down_proj",
"model.layers.19.mlp.down_proj",
"model.layers.21.mlp.down_proj",
"model.layers.22.mlp.down_proj",
.
.
.
]
"""
```
To reproduce the vLLM error:
```bash
vllm serve nm-testing/eldar-test
```
With this PR
------------
Models are loaded correctly without any errors.
**[Guided decoding performance optimization]** Sending the guided
decoding bitmask in xgrammar to the GPU
(`self.token_bitmask.to(scores.device)`) is a blocking operation that
prevents the CPU from pre-launching the sampler kernels. The CPU waits
until decode is complete, then copies the bitmask over. This PR changes
the operation to async via setting `non-blocking=True`.
(Current) The CPU is blocked on a `cudaStreamSynchronize` and only
pre-empts the sampling kernels after bitmask application. Below is the
Nsys profile for one decode phase from Llama 3.1 8B.
With the optimization, this is no longer the case:
---------
Signed-off-by: Ryan N <ryan.nguyen@centml.ai>
