streamft 0.6.0__tar.gz → 0.6.1__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (30) hide show
  1. {streamft-0.6.0 → streamft-0.6.1}/PKG-INFO +18 -7
  2. {streamft-0.6.0 → streamft-0.6.1}/README.md +17 -6
  3. {streamft-0.6.0 → streamft-0.6.1}/pyproject.toml +1 -1
  4. {streamft-0.6.0 → streamft-0.6.1}/streamft/adapter.py +14 -3
  5. {streamft-0.6.0 → streamft-0.6.1}/streamft/lora.py +23 -6
  6. {streamft-0.6.0 → streamft-0.6.1}/streamft/serve.py +9 -2
  7. {streamft-0.6.0 → streamft-0.6.1}/streamft.egg-info/PKG-INFO +18 -7
  8. {streamft-0.6.0 → streamft-0.6.1}/tests/test_adapter_quant.py +32 -0
  9. {streamft-0.6.0 → streamft-0.6.1}/LICENSE +0 -0
  10. {streamft-0.6.0 → streamft-0.6.1}/setup.cfg +0 -0
  11. {streamft-0.6.0 → streamft-0.6.1}/streamft/__init__.py +0 -0
  12. {streamft-0.6.0 → streamft-0.6.1}/streamft/__main__.py +0 -0
  13. {streamft-0.6.0 → streamft-0.6.1}/streamft/autotune.py +0 -0
  14. {streamft-0.6.0 → streamft-0.6.1}/streamft/cli.py +0 -0
  15. {streamft-0.6.0 → streamft-0.6.1}/streamft/data.py +0 -0
  16. {streamft-0.6.0 → streamft-0.6.1}/streamft/disk_quant_store.py +0 -0
  17. {streamft-0.6.0 → streamft-0.6.1}/streamft/doctor.py +0 -0
  18. {streamft-0.6.0 → streamft-0.6.1}/streamft/easy.py +0 -0
  19. {streamft-0.6.0 → streamft-0.6.1}/streamft/evals.py +0 -0
  20. {streamft-0.6.0 → streamft-0.6.1}/streamft/fit.py +0 -0
  21. {streamft-0.6.0 → streamft-0.6.1}/streamft/merge.py +0 -0
  22. {streamft-0.6.0 → streamft-0.6.1}/streamft/metrics.py +0 -0
  23. {streamft-0.6.0 → streamft-0.6.1}/streamft/stream_quant_lora.py +0 -0
  24. {streamft-0.6.0 → streamft-0.6.1}/streamft/stream_train.py +0 -0
  25. {streamft-0.6.0 → streamft-0.6.1}/streamft.egg-info/SOURCES.txt +0 -0
  26. {streamft-0.6.0 → streamft-0.6.1}/streamft.egg-info/dependency_links.txt +0 -0
  27. {streamft-0.6.0 → streamft-0.6.1}/streamft.egg-info/entry_points.txt +0 -0
  28. {streamft-0.6.0 → streamft-0.6.1}/streamft.egg-info/requires.txt +0 -0
  29. {streamft-0.6.0 → streamft-0.6.1}/streamft.egg-info/top_level.txt +0 -0
  30. {streamft-0.6.0 → streamft-0.6.1}/tests/test_streamft_smoke.py +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: streamft
3
- Version: 0.6.0
3
+ Version: 0.6.1
4
4
  Summary: Fine-tune models that do not fit: stream NF4 weights from disk, one transformer block at a time.
5
5
  Author-email: Gokhan Cava <byjameson@users.noreply.github.com>
6
6
  License: Apache-2.0
@@ -150,8 +150,9 @@ streamft train --model-dir checkpoints/qwen7b_nf4 --sft --steps 1000 \
150
150
  - On the reference Ampere card fp16 is slightly faster than bf16 at 7B (1.79 vs 1.91 s/step,
151
151
  same 5.36 GB peak); on Pascal it is the only option.
152
152
  - On a **4 GB** card use `--blocks-on-gpu 1`, `--seq 512` and `--vram-hard-gb 3.4`.
153
- Those rows are not measured yet — if you run one, post the JSON row and it can become a
154
- measured entry in `streamft fit`.
153
+ Those rows are not measured yet — if you run one, open a
154
+ [hardware report](.github/ISSUE_TEMPLATE/hardware-report.md) and it can become a measured
155
+ entry in `streamft fit`.
155
156
  - `streamft fit` prints this hint automatically when it sees a pre-Ampere GPU.
156
157
 
157
158
  ### Tokenizer cost and the diffusion question
@@ -210,12 +211,22 @@ Measured on the reference machine (Qwen2.5-3B, 60-step Dolly LoRA):
210
211
  | 8-bit | 8.56 MB | 0.0222 |
211
212
  | **4-bit** | **4.88 MB** | **0.0126** |
212
213
 
213
- 110 held-out Dolly rows, masked SFT loss: base 2.3197, original 2.3084, 8-bit 2.3075,
214
- 4-bit 2.3097 (+0.0013). The adapter loads in 0.14 s. The LoRA path is ~32% slower than the
215
- base on this card (8.19 vs 11.97 tok/s); merge it into a bf16 base for zero per-token
216
- overhead, or accept the LoRA path when storage/switching matters more than throughput.
214
+ 870 held-out Dolly rows, masked SFT loss: base 2.3752, original 2.3403, 8-bit 2.3401
215
+ (p = 0.81, no measurable loss), 4-bit 2.3420 (+0.0018, p = 0.001, 5% of the gain). A
216
+ second task, SST-2 (2,000 train / 400 test), goes from **71.25%** base to **93.75%**, and
217
+ both packed adapters stay at **93.75%** (masked loss 0.7940/0.7942 vs 0.7930; p = 0.50 /
218
+ 0.41). A third task, AG News (4-way), goes from 2.50% with the bare prompt (not followed)
219
+ to **62.25%**; 8-bit and 4-bit stay at **62.00%**. The adapter loads in 0.115 s. The LoRA
220
+ path is ~26% slower than the base on this card (8.75 vs 11.79 tok/s; a bf16 LoRA path
221
+ reaches 9.13); merge it into a bf16 base for zero per-token overhead, or accept the LoRA
222
+ path when storage/switching matters more than throughput.
217
223
  Rank reduction is not viable: SVD to rank 8 loses 35.8% of the update.
218
224
 
225
+ **A prompt can be the better rung.** On the same base, a 441-byte 4-shot SST-2 prompt
226
+ scores 94.0% (adapter 93.75%) and a 242-byte 4-shot AG News prompt scores 61.75% (adapter
227
+ 62.25%). If a few examples express the task, use the prompt; the adapter is for behaviour
228
+ a prompt cannot express, or when prompt tokens/context cost more than the storage.
229
+
219
230
  ### Measured on
220
231
 
221
232
  Every speed number here comes from **one machine**: RTX 3050 Laptop 6 GB, Windows 11,
@@ -120,8 +120,9 @@ streamft train --model-dir checkpoints/qwen7b_nf4 --sft --steps 1000 \
120
120
  - On the reference Ampere card fp16 is slightly faster than bf16 at 7B (1.79 vs 1.91 s/step,
121
121
  same 5.36 GB peak); on Pascal it is the only option.
122
122
  - On a **4 GB** card use `--blocks-on-gpu 1`, `--seq 512` and `--vram-hard-gb 3.4`.
123
- Those rows are not measured yet — if you run one, post the JSON row and it can become a
124
- measured entry in `streamft fit`.
123
+ Those rows are not measured yet — if you run one, open a
124
+ [hardware report](.github/ISSUE_TEMPLATE/hardware-report.md) and it can become a measured
125
+ entry in `streamft fit`.
125
126
  - `streamft fit` prints this hint automatically when it sees a pre-Ampere GPU.
126
127
 
127
128
  ### Tokenizer cost and the diffusion question
@@ -180,12 +181,22 @@ Measured on the reference machine (Qwen2.5-3B, 60-step Dolly LoRA):
180
181
  | 8-bit | 8.56 MB | 0.0222 |
181
182
  | **4-bit** | **4.88 MB** | **0.0126** |
182
183
 
183
- 110 held-out Dolly rows, masked SFT loss: base 2.3197, original 2.3084, 8-bit 2.3075,
184
- 4-bit 2.3097 (+0.0013). The adapter loads in 0.14 s. The LoRA path is ~32% slower than the
185
- base on this card (8.19 vs 11.97 tok/s); merge it into a bf16 base for zero per-token
186
- overhead, or accept the LoRA path when storage/switching matters more than throughput.
184
+ 870 held-out Dolly rows, masked SFT loss: base 2.3752, original 2.3403, 8-bit 2.3401
185
+ (p = 0.81, no measurable loss), 4-bit 2.3420 (+0.0018, p = 0.001, 5% of the gain). A
186
+ second task, SST-2 (2,000 train / 400 test), goes from **71.25%** base to **93.75%**, and
187
+ both packed adapters stay at **93.75%** (masked loss 0.7940/0.7942 vs 0.7930; p = 0.50 /
188
+ 0.41). A third task, AG News (4-way), goes from 2.50% with the bare prompt (not followed)
189
+ to **62.25%**; 8-bit and 4-bit stay at **62.00%**. The adapter loads in 0.115 s. The LoRA
190
+ path is ~26% slower than the base on this card (8.75 vs 11.79 tok/s; a bf16 LoRA path
191
+ reaches 9.13); merge it into a bf16 base for zero per-token overhead, or accept the LoRA
192
+ path when storage/switching matters more than throughput.
187
193
  Rank reduction is not viable: SVD to rank 8 loses 35.8% of the update.
188
194
 
195
+ **A prompt can be the better rung.** On the same base, a 441-byte 4-shot SST-2 prompt
196
+ scores 94.0% (adapter 93.75%) and a 242-byte 4-shot AG News prompt scores 61.75% (adapter
197
+ 62.25%). If a few examples express the task, use the prompt; the adapter is for behaviour
198
+ a prompt cannot express, or when prompt tokens/context cost more than the storage.
199
+
189
200
  ### Measured on
190
201
 
191
202
  Every speed number here comes from **one machine**: RTX 3050 Laptop 6 GB, Windows 11,
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
4
4
 
5
5
  [project]
6
6
  name = "streamft"
7
- version = "0.6.0"
7
+ version = "0.6.1"
8
8
  description = "Fine-tune models that do not fit: stream NF4 weights from disk, one transformer block at a time."
9
9
  readme = "README.md"
10
10
  requires-python = ">=3.10"
@@ -112,13 +112,23 @@ def tensor_error(a: torch.Tensor, b: torch.Tensor) -> float:
112
112
 
113
113
  def main(argv=None) -> int:
114
114
  ap = argparse.ArgumentParser(prog="streamft adapter")
115
- ap.add_argument("--adapter", required=True, help="plain adapter .pt (A/B tensors)")
116
- ap.add_argument("--bits", type=int, choices=[8, 4, 2], default=4)
115
+ ap.add_argument("--adapter", required=True,
116
+ help="adapter .pt, or a save dir containing adapter.pt")
117
+ ap.add_argument("--bits", type=int, choices=[8, 4, 2], default=4,
118
+ help="8=safe, 4=default, 2=not behaviourally validated")
117
119
  ap.add_argument("--group", type=int, default=64)
118
120
  ap.add_argument("--out", required=True)
119
121
  args = ap.parse_args(argv)
120
- state, meta = load_adapter(args.adapter)
121
122
  src = Path(args.adapter)
123
+ if src.is_dir():
124
+ src = src / "adapter.pt"
125
+ state, meta = load_adapter(src)
126
+ if not state:
127
+ raise SystemExit(f"no tensors in adapter: {src}")
128
+ lora_keys = {k for k in state if k.endswith(".A") or k.endswith(".B")}
129
+ if not lora_keys:
130
+ raise SystemExit(f"no LoRA A/B tensors in {src}; nothing to pack")
131
+ skipped = sorted(set(state) - lora_keys)
122
132
  out = Path(args.out)
123
133
  obj = save_quantized(state, out, args.bits, args.group, meta)
124
134
  back = dequantize_adapter(obj)
@@ -131,6 +141,7 @@ def main(argv=None) -> int:
131
141
  "compression": round(src.stat().st_size / max(out.stat().st_size, 1), 2),
132
142
  "max_relative_error": round(max(errs) if errs else 0.0, 4),
133
143
  "tensors": len(back),
144
+ "skipped_tensors": skipped,
134
145
  }, indent=1))
135
146
  return 0
136
147
 
@@ -32,17 +32,34 @@ class LoRALinear(nn.Module):
32
32
  dtype=torch.float32) / math.sqrt(base.in_features))
33
33
  self.B = nn.Parameter(torch.zeros(base.out_features, r, device=w.device, dtype=torch.float32))
34
34
  self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
35
+ self.fast_dtype = None # inference-only bf16/fp16 path; training stays fp32
35
36
 
36
37
  def forward(self, x):
37
38
  y = self.base(x)
38
- h = self.dropout(x).float()
39
- # A/B live on CPU (grads are accumulated there and the optimizer steps
40
- # there); hop them to the activation's device for the matmul. The low-rank
41
- # path is tiny, so this costs a negligible amount of PCIe traffic.
42
- A = self.A if self.A.device == h.device else self.A.to(h.device)
43
- B = self.B if self.B.device == h.device else self.B.to(h.device)
39
+ if self.fast_dtype is not None:
40
+ h = x.to(self.fast_dtype)
41
+ A = self.A.to(device=h.device, dtype=self.fast_dtype)
42
+ B = self.B.to(device=h.device, dtype=self.fast_dtype)
43
+ else:
44
+ h = self.dropout(x).float()
45
+ # A/B live on CPU (grads are accumulated there and the optimizer steps
46
+ # there); hop them to the activation's device for the matmul. The low-rank
47
+ # path is tiny, so this costs a negligible amount of PCIe traffic.
48
+ A = self.A if self.A.device == h.device else self.A.to(h.device)
49
+ B = self.B if self.B.device == h.device else self.B.to(h.device)
44
50
  return y + (self.scale * (h @ A.t() @ B.t())).to(y.dtype)
45
51
 
52
+ @torch.no_grad()
53
+ def set_fast(self, dtype=torch.bfloat16):
54
+ """Inference-only: run the low-rank path in `dtype` instead of fp32.
55
+
56
+ The adapter is trained in fp32 (small values would vanish in fp16); this is a
57
+ deployment switch. The weights are rounded once, at load time.
58
+ """
59
+ self.A.data = self.A.data.to(dtype)
60
+ self.B.data = self.B.data.to(dtype)
61
+ self.fast_dtype = dtype
62
+
46
63
  @torch.no_grad()
47
64
  def merge(self):
48
65
  self.base.weight.data += (self.scale * (self.B @ self.A)).to(self.base.weight.dtype)
@@ -108,7 +108,8 @@ def main(argv=None) -> int:
108
108
 
109
109
  ap = argparse.ArgumentParser(prog="streamft serve")
110
110
  ap.add_argument("--model", required=True, help="merged model dir or HF id")
111
- ap.add_argument("--adapter", default="", help="optional adapter dir with adapter.pt")
111
+ ap.add_argument("--adapter", default="",
112
+ help="optional adapter .pt, or a save dir containing adapter.pt")
112
113
  ap.add_argument("--host", default="0.0.0.0")
113
114
  ap.add_argument("--port", type=int, default=8000)
114
115
  ap.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
@@ -122,7 +123,7 @@ def main(argv=None) -> int:
122
123
  from pathlib import Path
123
124
 
124
125
  from .adapter import load_adapter
125
- from .lora import apply_lora, set_lora_state_dict
126
+ from .lora import LoRALinear, apply_lora, set_lora_state_dict
126
127
 
127
128
  p = Path(args.adapter)
128
129
  p = p / "adapter.pt" if p.is_dir() else p
@@ -130,6 +131,12 @@ def main(argv=None) -> int:
130
131
  r = int(next(v for k, v in state.items() if k.endswith(".A")).shape[0])
131
132
  apply_lora(model, r_attn=r, r_mlp=0)
132
133
  set_lora_state_dict(model, state)
134
+ if dtype != torch.float32:
135
+ # Inference-only bf16 low-rank path: ~4% faster than fp32 on the reference
136
+ # card and it matches the base dtype; training still uses fp32.
137
+ for m in model.modules():
138
+ if isinstance(m, LoRALinear):
139
+ m.set_fast(dtype)
133
140
  print(f"[serve] adapter {p.name} (r={r})", flush=True)
134
141
 
135
142
  _state.update(model=model, tok=tok, device=args.device, model_id=args.model)
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: streamft
3
- Version: 0.6.0
3
+ Version: 0.6.1
4
4
  Summary: Fine-tune models that do not fit: stream NF4 weights from disk, one transformer block at a time.
5
5
  Author-email: Gokhan Cava <byjameson@users.noreply.github.com>
6
6
  License: Apache-2.0
@@ -150,8 +150,9 @@ streamft train --model-dir checkpoints/qwen7b_nf4 --sft --steps 1000 \
150
150
  - On the reference Ampere card fp16 is slightly faster than bf16 at 7B (1.79 vs 1.91 s/step,
151
151
  same 5.36 GB peak); on Pascal it is the only option.
152
152
  - On a **4 GB** card use `--blocks-on-gpu 1`, `--seq 512` and `--vram-hard-gb 3.4`.
153
- Those rows are not measured yet — if you run one, post the JSON row and it can become a
154
- measured entry in `streamft fit`.
153
+ Those rows are not measured yet — if you run one, open a
154
+ [hardware report](.github/ISSUE_TEMPLATE/hardware-report.md) and it can become a measured
155
+ entry in `streamft fit`.
155
156
  - `streamft fit` prints this hint automatically when it sees a pre-Ampere GPU.
156
157
 
157
158
  ### Tokenizer cost and the diffusion question
@@ -210,12 +211,22 @@ Measured on the reference machine (Qwen2.5-3B, 60-step Dolly LoRA):
210
211
  | 8-bit | 8.56 MB | 0.0222 |
211
212
  | **4-bit** | **4.88 MB** | **0.0126** |
212
213
 
213
- 110 held-out Dolly rows, masked SFT loss: base 2.3197, original 2.3084, 8-bit 2.3075,
214
- 4-bit 2.3097 (+0.0013). The adapter loads in 0.14 s. The LoRA path is ~32% slower than the
215
- base on this card (8.19 vs 11.97 tok/s); merge it into a bf16 base for zero per-token
216
- overhead, or accept the LoRA path when storage/switching matters more than throughput.
214
+ 870 held-out Dolly rows, masked SFT loss: base 2.3752, original 2.3403, 8-bit 2.3401
215
+ (p = 0.81, no measurable loss), 4-bit 2.3420 (+0.0018, p = 0.001, 5% of the gain). A
216
+ second task, SST-2 (2,000 train / 400 test), goes from **71.25%** base to **93.75%**, and
217
+ both packed adapters stay at **93.75%** (masked loss 0.7940/0.7942 vs 0.7930; p = 0.50 /
218
+ 0.41). A third task, AG News (4-way), goes from 2.50% with the bare prompt (not followed)
219
+ to **62.25%**; 8-bit and 4-bit stay at **62.00%**. The adapter loads in 0.115 s. The LoRA
220
+ path is ~26% slower than the base on this card (8.75 vs 11.79 tok/s; a bf16 LoRA path
221
+ reaches 9.13); merge it into a bf16 base for zero per-token overhead, or accept the LoRA
222
+ path when storage/switching matters more than throughput.
217
223
  Rank reduction is not viable: SVD to rank 8 loses 35.8% of the update.
218
224
 
225
+ **A prompt can be the better rung.** On the same base, a 441-byte 4-shot SST-2 prompt
226
+ scores 94.0% (adapter 93.75%) and a 242-byte 4-shot AG News prompt scores 61.75% (adapter
227
+ 62.25%). If a few examples express the task, use the prompt; the adapter is for behaviour
228
+ a prompt cannot express, or when prompt tokens/context cost more than the storage.
229
+
219
230
  ### Measured on
220
231
 
221
232
  Every speed number here comes from **one machine**: RTX 3050 Laptop 6 GB, Windows 11,
@@ -50,3 +50,35 @@ def test_lora_linear_shapes():
50
50
  assert m.A.shape == (2, 8)
51
51
  assert m.B.shape == (4, 2)
52
52
  assert m(torch.randn(3, 8)).shape == (3, 4)
53
+
54
+
55
+ def test_cli_accepts_save_dir(tmp_path):
56
+ from streamft.adapter import main
57
+
58
+ save_dir = tmp_path / "run"
59
+ save_dir.mkdir()
60
+ torch.save(_state(), save_dir / "adapter.pt")
61
+ out = tmp_path / "adapter_q8.pt"
62
+ assert main(["--adapter", str(save_dir), "--bits", "8", "--out", str(out)]) == 0
63
+ assert out.exists()
64
+ back, _ = load_adapter(out)
65
+ assert set(back) == set(_state())
66
+
67
+
68
+ def test_cli_rejects_non_lora_state(tmp_path):
69
+ import pytest
70
+
71
+ from streamft.adapter import main
72
+
73
+ path = tmp_path / "plain.pt"
74
+ torch.save({"weight": torch.zeros(4, 4)}, path)
75
+ with pytest.raises(SystemExit, match="no LoRA"):
76
+ main(["--adapter", str(path), "--bits", "8", "--out", str(tmp_path / "q.pt")])
77
+
78
+
79
+ def test_lora_fast_path():
80
+ m = LoRALinear(nn.Linear(8, 4).to(torch.bfloat16), r=2)
81
+ m.set_fast(torch.bfloat16)
82
+ out = m(torch.randn(3, 8, dtype=torch.bfloat16))
83
+ assert out.shape == (3, 4)
84
+ assert out.dtype == torch.bfloat16
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes