feat: tests pass & can execute llama2

Author: Diogo-V

test/sparsity/test_marlin.py

@@ -69,6 +69,8 @@ def test_pack_unpack_equivalence(self):
             w_24, scales, zeros, n_bit=4, groupsize=group_size
         )
 
+        scales = scales.reshape(-1, w_q_24.shape[1])
+
         # Test pack/unpack equivalence
         q_w_comp, packed_scales, meta = pack_to_marlin_24(
             w_q_24, scales, num_bits, group_size

torchao/dtypes/affine_quantized_tensor.py

@@ -618,7 +618,9 @@ def from_plain(
 
         # Linear layers are (in_features, out_features) but the int_data that is reaching this point
         # is (out_features, in_features). We need to transpose it to match the expected shape in the marlin code.
+        # NOTE(reviewers): Please check if this is what I should do.
         q_w_24 = int_data.t()
+        scale = scale.reshape(-1, q_w_24.shape[1])
 
         if q_w_24.dtype != torch.int32:
             raise ValueError("Only `torch.int32` weights are supported.")
@@ -631,15 +633,14 @@ def from_plain(
 
         # NOTE: The current marlin 2:4 kernel supports both 4 and 8 bits quantization but fp8
         # will require a bit more work to get our current quantization flow to work with it.
-        # Check the below link for a reference:
-        # https://github.com/neuralmagic/nm-vllm/tree/main
+        # Check the link for a reference: https://github.com/neuralmagic/nm-vllm/tree/main
         num_bits = 4 if torch.max(q_w_24) < 16 else -1
         if num_bits not in [4]:
             raise ValueError(
-                f"Only {const.SUPPORTED_NUM_BITS} bits are supported, got {num_bits}."
+                f"Only {[4]} bits are supported, got {num_bits}."
             )
 
-        group_size = in_features // scale.shape[-1]
+        group_size = in_features // scale.shape[0]
         if group_size == 0:
             group_size = in_features
         assert group_size <= in_features, "Group size must be less than or equal to in_features."
@@ -1043,27 +1044,44 @@ def _linear_fp_act_int4_weight_sparse_marlin_check(input_tensor, weight_tensor,
         isinstance(weight_tensor.layout_type, MarlinSparseLayoutType)
     )
 
-
 def _linear_fp_act_int4_weight_sparse_marlin_impl(input_tensor, weight_tensor, bias):
-    from torchao.sparsity.marlin import marlin_24_workspace
+    from torchao.sparsity.marlin import marlin_24_workspace, const
 
     sparse_w_int4 = weight_tensor.layout_tensor.int_data
     scale = weight_tensor.layout_tensor.scale
     meta = weight_tensor.layout_tensor.meta
     original_shape = weight_tensor.layout_tensor.original_shape
     num_bits = weight_tensor.layout_tensor.num_bits
 
+    # Saves batch size for reshaping back to original shape after the matmul
+    # Reshapes tensor to (m, k) where m is in_features * batch and k is out_features
+    # NOTE(reviewers): Please check if I am handling the batch size correctly
+    batch_size = -1
+    if input_tensor.dim() == 3:
+        batch_size = input_tensor.size(0)
+        input_tensor = input_tensor.reshape(-1, input_tensor.shape[-1]).contiguous()
+
     size_m = input_tensor.shape[0]
-    size_n = original_shape[0]
+    size_n = original_shape[1]
     size_k = input_tensor.shape[1]
     workspace_24 = marlin_24_workspace(original_shape[1])
 
+    # Pad input_tensor dim 1 to a multiple of the marlin tile size (16)
+    if size_k % const.TILE != 0:
+        pad_size = find_multiple(size_k, const.TILE)
+        input_tensor = torch.nn.functional.pad(input_tensor, (0, pad_size - size_k))
+        size_k = pad_size
+
     out = torchao.ops.marlin_24_gemm(
         input_tensor, sparse_w_int4, meta, scale, 
         workspace_24, num_bits, size_m, size_n, size_k
     )
     torch.cuda.synchronize()
 
+    # Reshape back to original shape
+    if batch_size != -1:
+        out = out.reshape(batch_size, -1, out.shape[-1])
+
     if bias is not None:
         out += bias.to(out.dtype)
     return out

torchao/sparsity/marlin/__init__.py

@@ -333,7 +333,7 @@ def _from_marlin_scale(
     if group_size < size_k and group_size != -1:
         reverse_perms = reverse_marlin_24_scale_perm[num_bits]
         scales = scales.reshape((-1, len(reverse_perms)))[:, reverse_perms]
-        return scales.reshape((size_n, -1))
+        return scales.reshape((size_k // group_size, size_n))
     else:
         reverse_perms = reverse_marlin_24_scale_perm_single[num_bits]
         scales = scales.reshape((-1, len(reverse_perms)))[:, reverse_perms]

wip_test_llama2.py

@@ -0,0 +1,24 @@
+import torch
+from torchao import quantize_
+from torchao.quantization import int4_weight_only
+from torchao.dtypes import MarlinSparseLayoutType
+from transformers import AutoTokenizer, LlamaForCausalLM
+
+device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+name = "meta-llama/Llama-2-7b-hf"
+token = "your token"
+
+model = LlamaForCausalLM.from_pretrained(name, torch_dtype=torch.float16, token=token).to(device)
+tokenizer = AutoTokenizer.from_pretrained(name, token=token)
+
+prompt = "Hey, are you conscious? Can you talk to me? I'm"
+inputs = tokenizer(prompt, return_tensors="pt")
+
+# Quantize
+quantize_(model, int4_weight_only(layout_type=MarlinSparseLayoutType()))
+
+# Generate
+ids = inputs.input_ids.to(device)
+generate_ids = model.generate(ids, max_length=30)
+out = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+print(out)