16f7feee0a
The Cookbook's manual hardware simulator ("what if I had this setup") let users
pick a backend, but _apply_manual_hardware only accepted cuda/rocm/cpu_x86/
cpu_arm and silently coerced anything else to cuda. So selecting Apple/Metal
simulated a CUDA box instead — and ranked safetensors-only repos a Mac can't
serve, even though the rest of hwfit (services.hwfit.fit, the serve-command
generation) already supports Metal as GGUF-only via llama.cpp/Ollama.
Add "metal" to the accepted backends (now a named _MANUAL_BACKENDS set, kept a
subset of what fit.py understands) and set unified_memory=True for it — Apple
Silicon shares one memory pool with the GPU — while clearing that flag for the
discrete (cuda/rocm) and CPU backends. _apply_manual_hardware is lifted to
module scope so it is directly unit-testable; both route call sites are
unchanged.
Adds tests/test_hwfit_manual_backend.py, including an end-to-end check that a
simulated Metal box only recommends GGUF-servable models.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
288 lines
14 KiB
Python
288 lines
14 KiB
Python
import re
|
||
from copy import deepcopy
|
||
|
||
from fastapi import APIRouter
|
||
|
||
|
||
# Backends the manual hardware simulator accepts. Must stay a subset of what
|
||
# services.hwfit.fit understands so a simulated box ranks like a real one:
|
||
# "metal" routes through the Apple-Silicon path (GGUF-only, llama.cpp/Ollama),
|
||
# the CPU backends through the RAM/offload path, cuda/rocm through vLLM.
|
||
_MANUAL_BACKENDS = {"cuda", "rocm", "metal", "cpu_x86", "cpu_arm"}
|
||
|
||
|
||
def _apply_manual_hardware(system, manual_mode="", manual_gpu_count="", manual_vram_gb="", manual_ram_gb="", manual_backend=""):
|
||
"""Manual hardware is a "what if I had this setup" simulator —
|
||
REPLACES the detected hardware entirely instead of adding to it.
|
||
|
||
The previous additive behavior averaged the manual VRAM across
|
||
all GPUs (base + manual), which meant adding "1× 400 GB" on top
|
||
of "2× 70 GB" only nudged the per-GPU cap from 70 to 180 GB
|
||
(= 540 / 3), so GGUF models bigger than that still didn't surface
|
||
— exactly the "cap stuck at detected level" bug the user hit.
|
||
"""
|
||
manual_mode = (manual_mode or "").lower()
|
||
if manual_mode not in {"gpu", "ram"}:
|
||
return system
|
||
|
||
try:
|
||
override_ram_gb = float(manual_ram_gb) if manual_ram_gb else 0
|
||
except ValueError:
|
||
override_ram_gb = 0
|
||
override_ram_gb = max(0.0, override_ram_gb)
|
||
if override_ram_gb:
|
||
# Replace RAM, don't add. The number in the field is the
|
||
# TOTAL system memory the user wants to simulate.
|
||
system["available_ram_gb"] = round(override_ram_gb, 1)
|
||
system["total_ram_gb"] = round(override_ram_gb, 1)
|
||
system["manual_hardware"] = True
|
||
|
||
if manual_mode == "ram":
|
||
# RAM-only simulation — wipe GPU entirely so the ranker uses
|
||
# CPU/RAM paths.
|
||
system["has_gpu"] = False
|
||
system["gpu_name"] = None
|
||
system["gpu_vram_gb"] = 0
|
||
system["gpu_count"] = 0
|
||
system["gpus"] = []
|
||
system["gpu_groups"] = []
|
||
system["backend"] = "cpu_x86"
|
||
system.pop("unified_memory", None)
|
||
return system
|
||
|
||
try:
|
||
count = int(manual_gpu_count) if manual_gpu_count else 1
|
||
except ValueError:
|
||
count = 1
|
||
try:
|
||
vram_each = float(manual_vram_gb) if manual_vram_gb else 8.0
|
||
except ValueError:
|
||
vram_each = 8.0
|
||
count = max(1, min(count, 16))
|
||
vram_each = max(1.0, vram_each)
|
||
backend = (manual_backend or system.get("backend") or "cuda").lower()
|
||
if backend not in _MANUAL_BACKENDS:
|
||
backend = "cuda"
|
||
total_vram = round(vram_each * count, 1)
|
||
gpu_name = f"Simulated {backend.upper()} GPU" + (f" × {count}" if count > 1 else "")
|
||
system["has_gpu"] = True
|
||
system["gpu_name"] = gpu_name
|
||
system["gpu_vram_gb"] = total_vram
|
||
system["gpu_count"] = count
|
||
system["gpus"] = [
|
||
{"index": i, "name": gpu_name, "vram_gb": vram_each}
|
||
for i in range(count)
|
||
]
|
||
# Single homogeneous pool — vram_each here is the ACTUAL per-GPU
|
||
# VRAM the user entered, not an average. That's the whole point:
|
||
# raising vram_each lifts the per-GPU cap (GGUF, tensor-parallel
|
||
# math) all the way up, not just by a small fraction.
|
||
system["gpu_groups"] = [{
|
||
"name": gpu_name,
|
||
"vram_each": vram_each,
|
||
"count": count,
|
||
"indices": list(range(count)),
|
||
"vram_total": total_vram,
|
||
}]
|
||
system["homogeneous"] = True
|
||
system["backend"] = backend
|
||
# Apple Silicon shares one unified memory pool with the GPU; flag it so
|
||
# the API/UI report it the way real Metal detection does. Discrete GPUs
|
||
# (cuda/rocm) and the CPU backends carry separate VRAM, so clear any
|
||
# stale flag a previous detection left on the dict.
|
||
if backend == "metal":
|
||
system["unified_memory"] = True
|
||
else:
|
||
system.pop("unified_memory", None)
|
||
return system
|
||
|
||
|
||
def setup_hwfit_routes():
|
||
router = APIRouter(prefix="/api/hwfit", tags=["hwfit"])
|
||
|
||
@router.get("/system")
|
||
def get_system(host: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False):
|
||
"""Detect and return current system hardware info. Pass host=user@server for remote.
|
||
fresh=true bypasses the per-host cache (the Rescan button)."""
|
||
from services.hwfit.hardware import detect_system
|
||
return detect_system(host=host, ssh_port=ssh_port, platform=platform, fresh=fresh)
|
||
|
||
@router.get("/models")
|
||
def get_models(use_case: str = "", sort: str = "score", limit: int = 50, search: str = "", host: str = "", quant: str = "", ctx: str = "", gpu_count: str = "", gpu_group: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, manual_mode: str = "", manual_gpu_count: str = "", manual_vram_gb: str = "", manual_ram_gb: str = "", manual_backend: str = "", ignore_detected_gpu: bool = False, ignore_detected_ram: bool = False):
|
||
"""Rank LLM models against detected hardware and return scored results.
|
||
gpu_count: override GPU count (0 = CPU only, 1-N = simulate N GPUs of the
|
||
active group). gpu_group: index into system.gpu_groups (the homogeneous
|
||
pools) to target — empty/auto = the largest pool. vLLM can only
|
||
tensor-parallel across identical GPUs, so we never mix pools.
|
||
fresh=true bypasses the hardware-detection cache."""
|
||
from services.hwfit.hardware import detect_system
|
||
from services.hwfit.fit import rank_models
|
||
from services.hwfit.models import get_models, model_catalog_path
|
||
system = deepcopy(detect_system(host=host, ssh_port=ssh_port, platform=platform, fresh=fresh))
|
||
if system.get("error"):
|
||
return {"system": system, "models": [], "error": system["error"]}
|
||
if not get_models():
|
||
return {
|
||
"system": system,
|
||
"models": [],
|
||
"error": f"Model catalog missing or empty: {model_catalog_path()}",
|
||
}
|
||
|
||
if ignore_detected_gpu:
|
||
system["has_gpu"] = False
|
||
system["gpu_name"] = None
|
||
system["gpu_vram_gb"] = 0
|
||
system["gpu_count"] = 0
|
||
system["gpus"] = []
|
||
system["gpu_groups"] = []
|
||
if ignore_detected_ram:
|
||
system["available_ram_gb"] = 0
|
||
system["total_ram_gb"] = 0
|
||
|
||
system = _apply_manual_hardware(system, manual_mode, manual_gpu_count, manual_vram_gb, manual_ram_gb, manual_backend)
|
||
|
||
# Keep the raw detection around so the UI can still show the box's full
|
||
# GPU complement even while we rank against one homogeneous pool.
|
||
system["detected_gpu_vram_gb"] = system.get("gpu_vram_gb")
|
||
system["detected_gpu_count"] = system.get("gpu_count")
|
||
|
||
groups = system.get("gpu_groups") or []
|
||
# Resolve the target homogeneous pool. Default (auto) = the largest pool,
|
||
# which for a uniform box is simply "all the GPUs" — no behaviour change.
|
||
grp = None
|
||
if groups:
|
||
try:
|
||
gidx = int(gpu_group) if gpu_group != "" else 0
|
||
except ValueError:
|
||
gidx = 0
|
||
if 0 <= gidx < len(groups):
|
||
grp = groups[gidx]
|
||
|
||
def _apply_group(g, n):
|
||
n = max(1, min(n, g["count"]))
|
||
system["gpu_count"] = n
|
||
system["gpu_vram_gb"] = round(g["vram_each"] * n, 1)
|
||
system["gpu_name"] = g["name"]
|
||
system["active_group"] = {**g, "use_count": n}
|
||
|
||
if gpu_count != "":
|
||
n = int(gpu_count)
|
||
if n == 0:
|
||
# RAM-only mode: rank against system memory, offload allowed.
|
||
system["has_gpu"] = False
|
||
system["gpu_vram_gb"] = 0
|
||
system["gpu_count"] = 0
|
||
system["gpu_only"] = False
|
||
system.pop("active_group", None)
|
||
elif grp:
|
||
_apply_group(grp, n)
|
||
system["gpu_only"] = True
|
||
else:
|
||
# No per-GPU detail (older detection) — assume uniform split.
|
||
single_vram = (system.get("gpu_vram_gb") or 0) / (system.get("gpu_count") or 1)
|
||
system["gpu_count"] = max(1, n)
|
||
system["gpu_vram_gb"] = round(single_vram * max(1, n), 1)
|
||
system["gpu_only"] = True
|
||
elif grp:
|
||
# No explicit count, but we still pin to one pool so heterogeneous
|
||
# boxes rank against a real mixable group, not a fictional VRAM sum.
|
||
# gpu_only stays off here so the default view still surfaces offload.
|
||
_apply_group(grp, grp["count"])
|
||
|
||
try:
|
||
target_context = int(ctx) if ctx else None
|
||
except ValueError:
|
||
target_context = None
|
||
if target_context is not None:
|
||
target_context = max(1024, min(target_context, 1000000))
|
||
|
||
results = rank_models(system, use_case=use_case or None, limit=limit, search=search or None, sort=sort, quant=quant or None, target_context=target_context)
|
||
return {"system": system, "models": results}
|
||
|
||
@router.get("/profiles")
|
||
def get_serve_profiles(model: str = "", host: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, serve_weights_gb: float = 0.0, serve_quant: str = ""):
|
||
"""Compute llama.cpp serve profiles (Quality/Balanced/Speed) for `model`
|
||
against the detected hardware on `host` (or local). Returns concrete
|
||
flags (n_gpu_layers, n_cpu_moe, cache_type, ctx) the serve UI can apply.
|
||
|
||
`model` is matched against the catalog by name; if it's not in the
|
||
catalog (e.g. an ad-hoc HF repo), pass enough hints via a minimal synthetic
|
||
entry isn't possible here, so we return [] and the UI keeps manual flags.
|
||
"""
|
||
from services.hwfit.hardware import detect_system
|
||
from services.hwfit.models import get_models
|
||
from services.hwfit.profiles import compute_serve_profiles
|
||
system = detect_system(host=host, ssh_port=ssh_port, platform=platform, fresh=fresh)
|
||
if system.get("error"):
|
||
return {"system": system, "profiles": [], "error": system["error"]}
|
||
catalog = {m.get("name"): m for m in (get_models() or [])}
|
||
|
||
def _norm(s):
|
||
# Normalize for matching: drop org/ prefix, a trailing -GGUF/-gguf
|
||
# marker, and any quant tag, lowercase. So "DeepSeek-Coder-V2-Lite-
|
||
# Instruct-GGUF" (a local folder name) matches catalog entry
|
||
# "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct".
|
||
s = (s or "").lower().strip()
|
||
s = s.split("/")[-1] # drop org prefix
|
||
s = re.sub(r"[-_.]?gguf$", "", s) # drop trailing gguf marker
|
||
s = re.sub(r"[-_.](q\d[^/]*|iq\d[^/]*|fp8|bf16|f16|awq[^/]*|gptq[^/]*)$", "", s)
|
||
return s
|
||
|
||
m = catalog.get(model)
|
||
if m is None and model:
|
||
want = _norm(model)
|
||
for name, entry in catalog.items():
|
||
nn = _norm(name)
|
||
if nn and (nn == want or want.endswith(nn) or nn.endswith(want)):
|
||
m = entry
|
||
break
|
||
if m is None:
|
||
return {"system": system, "profiles": [], "error": "model not in catalog"}
|
||
# Surface the model's trained context limit so the serve UI can clamp a
|
||
# user-typed context down to it (asking for ctx > n_ctx_train overflows
|
||
# and, with a quantized KV cache, can crash the GPU).
|
||
model_ctx_max = 0
|
||
for k in ("context_length", "max_position_embeddings", "n_ctx_train", "context"):
|
||
v = m.get(k)
|
||
if isinstance(v, (int, float)) and v > 0:
|
||
model_ctx_max = int(v)
|
||
break
|
||
return {
|
||
"system": system,
|
||
"profiles": compute_serve_profiles(
|
||
system, m,
|
||
serve_weights_gb=(serve_weights_gb or None),
|
||
serve_quant=(serve_quant or None),
|
||
),
|
||
"model_ctx_max": model_ctx_max,
|
||
}
|
||
|
||
@router.get("/image-models")
|
||
def get_image_models(sort: str = "fit", search: str = "", host: str = "", gpu_count: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, manual_mode: str = "", manual_gpu_count: str = "", manual_vram_gb: str = "", manual_ram_gb: str = "", manual_backend: str = "", ignore_detected_gpu: bool = False, ignore_detected_ram: bool = False):
|
||
"""Rank image generation models against detected hardware."""
|
||
from services.hwfit.hardware import detect_system
|
||
from services.hwfit.image_models import rank_image_models
|
||
system = deepcopy(detect_system(host=host, ssh_port=ssh_port, platform=platform, fresh=fresh))
|
||
if system.get("error"):
|
||
return {"system": system, "models": [], "error": system["error"]}
|
||
if ignore_detected_gpu:
|
||
system["has_gpu"] = False
|
||
system["gpu_name"] = None
|
||
system["gpu_vram_gb"] = 0
|
||
system["gpu_count"] = 0
|
||
system["gpus"] = []
|
||
system["gpu_groups"] = []
|
||
if ignore_detected_ram:
|
||
system["available_ram_gb"] = 0
|
||
system["total_ram_gb"] = 0
|
||
system = _apply_manual_hardware(system, manual_mode, manual_gpu_count, manual_vram_gb, manual_ram_gb, manual_backend)
|
||
# Image models use a single GPU — always use per-GPU VRAM
|
||
gpu_vrams = [float(g.get("vram_gb") or 0) for g in (system.get("gpus") or []) if isinstance(g, dict)]
|
||
single_vram = max(gpu_vrams) if gpu_vrams else ((system.get("gpu_vram_gb") or 0) / max(system.get("gpu_count") or 1, 1))
|
||
system["gpu_vram_gb"] = single_vram
|
||
system["gpu_count"] = 1 if single_vram > 0 else 0
|
||
results = rank_image_models(system, search=search or None, sort=sort)
|
||
return {"system": system, "models": results}
|
||
|
||
return router
|